System and method for broadcasting messages to a virtual data processing accelerator
By using broadcast protocols and encryption technology between data processing accelerators, secure communication between virtual data processing accelerators is ensured, and the problem of insecure communication of data processing accelerators is solved, and the secure execution of tasks and data integrity is achieved.
Patent Information
- Application Number
- CN202011455706.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2019-12-10
- Filing Date
- 2020-12-10
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2040-12-10
AI Technical Summary
Communication between data processing accelerators is unsafe and vulnerable to malware or intrusions, resulting in data and code being stolen or tampered with, and failing to ensure the security and integrity of processing tasks.
Adopt the broadcast protocol to encrypt messages by responding to broadcast instructions, use a security key pair to generate a public key and a session key, ensure secure communication between virtual data processing accelerators, and use the encryption module and broadcast module to encrypt messages and keys. Propagate and decrypt the private key to ensure the security of the message.
It realizes secure communication between data processing accelerators, prevents data and code from being stolen or tampered, and ensures the secure execution of processing tasks and the confidentiality and integrity of data.
Smart Images

Figure CN112953717B_ABST
Abstract
Description
Technical Field
[0001] Embodiments of the present disclosure generally relate to artificial intelligence model training and inference. More specifically, embodiments of the present disclosure relate to systems and methods for broadcast protocols for securely broadcasting messages to data processing accelerators using virtual communication channels. Background Art
[0002] Data processing accelerators (DPAs) configured to communicate with a host computing device are typically unable to communicate securely with each other. Communication between DPAs, such as point-to-point communication, is useful so that two or more DPAs can collaborate and coordinate to perform data processing tasks on behalf of the host computing device. However, it is important that the DPAs communicate securely so that processing tasks performed by the collaborating DPAs are performed securely, so that another computing entity cannot change the results produced by the communicating DPA, and cannot steal code or data from any of the communicating DPAs. Summary of the invention
[0003] Embodiments of the present disclosure disclose a method and apparatus for broadcasting a message to one or more virtual data processing (DP) accelerators, and a data processing system.
[0004] In a first aspect of an embodiment of the present disclosure, a method for broadcasting a message to one or more virtual data processing (DP) accelerators is provided, comprising: in response to receiving a broadcast instruction from an application, encrypting the broadcast message based on a broadcast session key for a broadcast communication session, the broadcast instruction specifying one or more virtual DP accelerators among a plurality of virtual DP accelerators to receive the broadcast message; determining one or more public keys of one or more security key pairs, each of the security key pairs being associated with one of the specified virtual DP accelerators; encrypting the broadcast session key based on the determined one or more public keys; and broadcasting the encrypted broadcast message and the one or more encrypted broadcast session keys to adjacent virtual DP accelerators for dissemination, wherein the specified virtual DP accelerator decrypts the encrypted broadcast session key based on a corresponding private key associated with the specified virtual DP accelerator, wherein the message is decrypted based on the broadcast session key.
[0005] In a second aspect of an embodiment of the present disclosure, a device for broadcasting a message to one or more virtual data processing (DP) accelerators is provided, comprising: a first encryption module for encrypting a broadcast message based on a broadcast session key for a broadcast communication session in response to receiving a broadcast instruction from an application, wherein the broadcast instruction specifies one or more virtual DP accelerators among a plurality of virtual DP accelerators to receive the broadcast message; a determination module for determining one or more public keys of one or more security key pairs, each of the security key pairs being associated with one of the specified virtual DP accelerators; a second encryption module for encrypting the broadcast session key based on the determined one or more public keys; and a broadcast module for broadcasting the encrypted broadcast message and the one or more encrypted broadcast session keys to adjacent virtual DP accelerators for dissemination, wherein the specified virtual DP accelerator decrypts the encrypted broadcast session key based on a corresponding private key associated with the specified virtual DP accelerator, wherein the message is decrypted based on the broadcast session key.
[0006] In a third aspect of an embodiment of the present disclosure, a non-transitory machine-readable medium is provided, in which instructions are stored. When the instructions are executed by a processor, the processor executes the method as described in the first aspect of the embodiment of the present application.
[0007] In a fourth aspect of an embodiment of the present disclosure, a data processing system is provided, comprising: a processor; and a memory coupled to the processor to store instructions, which, when executed by the processor, causes the processor to execute the method described in the first aspect of the embodiment of the present application. BRIEF DESCRIPTION OF THE DRAWINGS
[0008] Embodiments of the present disclosure are illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings in which like references indicate similar elements.
[0009] Figure 1 is a block diagram illustrating a secure processing system according to one embodiment.
[0010] Figure 2A and 2B is a block diagram illustrating a secure computing environment between one or more hosts and one or more virtual data processing accelerators according to one embodiment.
[0011] Figure 3 is a block diagram illustrating a method in which a host and a virtual data processing accelerator or two data processing accelerators generate a session key for secure communication according to an embodiment.
[0012] Figure 4 is a block diagram illustrating a hardware configuration of a host computing device and multiple virtual data processing accelerators that securely communicate with each other according to an embodiment.
[0013] Figure 5 is a block diagram illustrating a secure communication adjacency table between a host device and a plurality of virtual data processing accelerators according to an embodiment.
[0014] Figure 6 is a block diagram illustrating a method 600 of a host device instructing a plurality of virtual data processing accelerators to configure themselves for secure communications according to an embodiment.
[0015] Figure 7 is a block diagram illustrating a method in which a virtual data processing accelerator configures itself for secure communications with one or more other data processing accelerators according to an embodiment.
[0016] Figure 8 is a block diagram illustrating a method in which a virtual data processing accelerator receives a processing task from a host and executes one or more subtasks of the task through one or more additional virtual data processing accelerators according to an embodiment.
[0017] Fig. 9 is a flow chart illustrating a broadcast protocol of a host and one or more virtual data processing accelerators according to an embodiment.
[0018] Fig.10 is a flow chart illustrating an example method for a host to perform a broadcast according to an embodiment.
[0019] Fig.11 is a flow chart illustrating an example method for a virtual data processing accelerator to perform broadcasting according to an embodiment.
[0020] Fig.12 is a block diagram illustrating a data processing system according to one embodiment. DETAILED DESCRIPTION
[0021] Various embodiments and aspects of the present disclosure will be described with reference to the details discussed below, and the accompanying drawings will illustrate various embodiments. The following description and the accompanying drawings are illustrative of the present disclosure and should not be construed as limiting the present disclosure. Many specific details are described to provide a comprehensive understanding of various embodiments of the present disclosure. However, in some cases, in order to provide a brief discussion of embodiments of the present disclosure, known or conventional details are not described.
[0022] References in the specification to "one embodiment" or "an embodiment" mean that a particular feature, structure, or characteristic described in conjunction with the embodiment may be included in at least one embodiment of the present disclosure. The phrase "in one embodiment" appearing in various places in the specification does not necessarily refer to the same embodiment.
[0023] The following embodiments relate to using data processing (DP) accelerators to increase the processing throughput of specific types of operations that can be offloaded (or delegated) from a host device to one or more DP accelerators. The DP accelerator may be a graphics processing unit (GPU), an artificial intelligence (AI) accelerator, a math coprocessor, a digital signal processor (DSP), or other type of processor. The DP accelerator may be a proprietary design, such as AI accelerator or another GPU, etc. Although the embodiments are shown and described using a host device securely coupled to one or more DP accelerators, the concepts described herein may be more generally implemented as a distributed processing system.
[0024] The host device and the DP accelerator may be interconnected via a high-speed bus, such as a peripheral component express interconnect (PCIe) or other high-speed bus. The DP accelerator may be configured to include a plurality of virtual DP (VDP) accelerators. Each of the VDP accelerators may communicate with the host via a virtual communication channel. Each of the VDP accelerators may communicate with at least one of the VDP accelerators via a different virtual communication channel. The virtual communication channels are each protected using a different session key for each channel.
[0025] The host device and each VDP accelerator may exchange keys and initiate a secure channel over a PCIe bus before performing the operations of the various aspects of the present invention described below. This document describes an embodiment for generating one or more keys for protecting communications between a host and a VDP accelerator and for protecting communications between any two VDP accelerators among a plurality of VDP accelerators. In an embodiment, communications between any two VDP accelerators use one or more keys that are unique relative to any other two VDP accelerators. Some of the operations performed by the VDP accelerator include the VDP accelerator using an artificial intelligence (AI) model to perform inference using data provided by the host device. Before delegating AI model inference to the VDP accelerator, a secure virtual communication channel is established between the host and the VDP accelerator and between the VDP accelerator and any other VDP accelerator that may participate in AI model inference.
[0026] Embodiments disclose systems and methods for broadcasting messages to one or more virtual data processing (DP) accelerators. In response to receiving a broadcast instruction from an application, the broadcast instruction specifies one or more virtual DP accelerators among multiple virtual DP accelerators to receive the broadcast message, and the system encrypts the broadcast message based on a broadcast session key for a broadcast communication session. The system determines one or more public keys of one or more security key pairs that are each associated with one of the specified virtual DP accelerators. The system encrypts the broadcast session key based on the determined one or more public keys. The system broadcasts the encrypted broadcast message and one or more encrypted broadcast session keys to adjacent virtual DP accelerators for dissemination, wherein the specified virtual DP accelerator decrypts the encrypted broadcast session key based on a corresponding private key associated with the specified virtual DP accelerator, wherein the message is decrypted based on the broadcast session key.
[0027] Any of the above functions can be programmed as executable instructions on one or more non-transitory computer-readable media. When the executable instructions are executed by a processing system having at least one hardware processor, the processing system causes the functions to be implemented.
[0028] Any of the above functions may be implemented by a processing system having at least one hardware processor coupled to a memory programmed with executable instructions that, when executed, cause the processing system to implement the functions.
[0029] Figure 1 is a block diagram illustrating an example of a system configuration for securing communications between a host 104 and data processing (DP) accelerators 105-107 according to some embodiments. Figure 1 , the system configuration 100 includes, but is not limited to, one or more client devices 101-102 communicatively coupled to a DP server 104 (e.g., a host) via a network 103. The client devices 101-102 may be any type of client device, such as a personal computer (e.g., a desktop, a laptop, and a tablet computer), a "thin" client, a personal digital assistant (PDA), a web-enabled appliance, a smart watch, or a mobile phone (e.g., a smartphone), etc. Alternatively, the client devices 101-102 may be other servers. The network 103 may be any type of network, such as a local area network (LAN), a wide area network (WAN) such as the Internet, or a combination thereof, wired or wireless.
[0030] The server (e.g., host) 104 can be any kind of server or server cluster, such as a web or cloud server, an application server, a backend server, or a combination thereof. The server 104 also includes an interface (not shown) to allow clients such as client devices 101-102 to access resources or services provided by the server 104 (such as resources and services provided by the DP accelerator via the server 104). For example, the server 104 can be a cloud server or a server in a data center, providing various cloud services to clients, such as cloud storage, cloud computing services, artificial intelligence training services, data mining services, etc. The server 104 can be configured as part of a software as a service (SaaS) or platform as a service (PaaS) system on a cloud, which can be a private cloud, a public cloud, or a hybrid cloud. The interface may include a web interface, an application programming interface (API), and / or a command line interface (CLI).
[0031] For example, the client, in this example, the user application of the client device 101 (e.g., a web browser, an application) may send or transmit instructions for execution (e.g., artificial intelligence (AI) training, inference instructions, etc.) to the server 104, and the instructions are received by the server 104 via the interface through the network 103. In response to the instructions, the server 104 communicates with the DP accelerators 105-107 to complete the execution of the instructions. In some embodiments, the instructions are machine learning type instructions, where the DP accelerator, as a dedicated machine or processor, can execute the instructions many times faster than the server 104. Therefore, the server 104 can control / manage the execution of one or more DP accelerators in a distributed manner. The server 104 then returns the execution results to the client devices 101-102. The DP accelerator or AI accelerator may include one or more dedicated processors, such as those available from Baidu, Inc. Artificial Intelligence (AI) Chipset, Inc. Or alternatively, the DP accelerator may be an AI chipset from another AI chipset provider.
[0032] According to one embodiment, each of the applications accessing any one of the DP accelerators 105-107 hosted by the data processing server 104 (also referred to as the host) can be verified to be provided by a trusted source or supplier. Each DP accelerator can be configured as multiple (e.g., four) virtual DP (VDP) accelerators. Thus, for example, the DP accelerator 105 can be configured as four virtual DP accelerators 105A, 105B, 105C, and 105D. Each of the applications can be opened and executed in a trusted execution environment (TEE) that is specially configured and executed by the central processing unit (CPU) of the host 104. When the application is configured to access a VDP accelerator, such as any one of 105A-105D, 106A-106D, or 107A-107D, an obfuscated connection can be established between the host 104 and the corresponding one of the VDP accelerators, so that the data exchanged between the host 104 and the VDP accelerator is protected from attacks from malware / intrusion.
[0033] Figure 2A 2 is a block diagram illustrating an example of a multi-layered protection solution for obfuscated communications between a host system 104 and a data processing (DP) accelerator 105-107 according to some embodiments. In one embodiment, the system 200 provides a protection solution for obfuscated communications between the host 104 and the DP accelerators 105-107 with or without hardware modifications to the DP accelerators. Figure 2A , the host machine or server 104 can be depicted as a system with one or more layers to be protected from intrusion, such as user applications 205, runtime libraries 206, drivers 209, operating system 211, and hardware 213 (e.g., security module (Trusted Platform Module (TPM) / Central Processing Unit (CPU)). A memory security application 207 can run in a sandboxed memory. Below the application 205 and runtime libraries 206, one or more drivers 209 can be installed to connect with the hardware 213 and / or DP accelerators 105-107.
[0034] The hardware 213 may include one or more processors 201 and a storage device 204. The storage device 204 may include one or more artificial intelligence (AI) models 202, and one or more kernels 203. The kernel 203 may include a signature kernel, a watermark-enabled kernel, an encryption and / or decryption kernel, etc. The signature kernel may digitally sign any input according to the programming of the kernel when executed. The watermark-enabled kernel may extract a watermark from a data object (e.g., an AI model or other data object). The watermark-enabled kernel may also implant a watermark into an AI model, an inference output, or other data object. A watermark kernel (e.g., a watermark-inheriting kernel) may inherit a watermark from another data object and implant the watermark into a different object, such as an inference output or an AI model. As used herein, a watermark is an identifier associated with an AI model or an inference generated by an AI model, and may be implanted into an AI model or an inference generated by an AI model. For example, a watermark may be implanted into one or more weight variables or bias variables. Alternatively, one or more nodes (e.g., a fake node that is not used or is unlikely to be used by an artificial intelligence model) may be created to implant or store a watermark.
[0035] The host machine 104 is typically a CPU system that can control and manage the execution of jobs on the host machine 104 and / or the DP accelerators 105-107. In order to protect / obfuscate the communication channel 215 between the DP accelerators 105-107 and the host machine 104, different components may be required to protect different layers of the host system that are susceptible to data intrusion or attack. For example, a trusted execution environment (TEE) can protect the user application 205 layer and the runtime library 206 layer from data intrusion. The communication channel 215 can be configured to carry multiple virtual channels. For example, the host 104 can communicate with the DP accelerator 105 through a single physical channel 215 including multiple, for example, four virtual communication channels 215A, 215B, 215C, and 215D. In one embodiment, the virtual communication channels 215A-215D can be time-division multiplexed on the physical channel communication channel 215. Each virtual communication channel can have a session key different from the session key of any other virtual communication channel for protecting communications on the virtual communication channel. In one embodiment, messages sent over the virtual communication channel may be sent over the physical communication channel on a queued basis using any of first-in-first-out, biggest-first-out, smallest-first-out, last-in-first-out, or other queue logic.
[0036] According to some embodiments, system 200 includes host system 104 and DP accelerators 105-107. DP accelerators may include The DP accelerator may be configured as a plurality of virtual DP accelerators according to the configuration instructions received from the host 104. In an embodiment, the configuration instructions may be provided to each DP accelerator at, for example, boot time, and the configuration is maintained until the next boot sequence of the host 104. In an embodiment, each virtual DP accelerator communicates with an adjacent host device 104 or other virtual DP accelerators via a virtual communication channel defined by points (nodes) communicating on the virtual communication channel. For example, the host 104 may communicate with the DP accelerator via a physical communication channel 215 defined by the physical port 0 of the host 104 to the physical port 0 of the DP accelerator 105. The host 104 may instruct the DP accelerator 105 to configure itself as four (4) virtual DP accelerators 105A, 105B, 105C, and 105D. The host 104 and the DP accelerator 105 may configure themselves to treat the communication channel 215 as four virtual channels 215A, 215B, 215C, and 215D. Thus, the host 104 communicates with the virtual DP accelerator (VDPA) 105A using the virtual communication channel 215A. Similarly, the host 104 communicates with the VPDA 105B via the virtual communication channel 215B, and similarly, the VPDAs 105C and 105D communicate via the virtual communication channels 215C and 215D. The host 104 may maintain an adjacency table that relates processes running on the host 104 to the physical communication ports of the host 104, the virtual communication ports of the physical communication ports, the corresponding virtual data processing accelerators, and session keys. In one embodiment, each session key is unique with respect to other session keys used for other virtual communication channels.
[0037] In one embodiment, the host system 104 includes hardware with one or more CPUs 213, which are equipped with a security module (such as a trusted platform module (TPM)) within the host machine 104. The TPM is a dedicated chip on an endpoint device that stores cryptographic keys (e.g., RSA cryptographic keys) dedicated to the host system for hardware authentication. Each TPM chip may include one or more RSA key pairs (e.g., public and private key pairs), called endorsement keys (EKs) or endorsement certificates (ECs), i.e., root keys. The key pairs are maintained inside the TPM chip and cannot be accessed by software. The EK or EC can then hash the critical parts of the firmware and software before they are executed to protect the system from unauthorized firmware and software modifications. The TPM chip on the host machine can therefore be used as a trusted root for secure boot. The TPM chip may include secure storage for storing, for example, keys that are rooted in hardware and keys derived from the rooted keys. In one embodiment, the secure storage may include a rooted asymmetric key pair (RK): a public key (PK_RK) and a private key (SK_RK) of the asymmetric rooted key (RK) pair.
[0038] The TPM chip also protects the driver 209 and the operating system (OS) 211 in the working kernel space to communicate with the DP accelerator 105-107. Here, the driver 209 is provided by the DP accelerator vendor and can be used as a driver for the user application to control the communication channel 215 between the host and the DP accelerator. Because the TPM chip and the secure boot processor protect the OS 211 and the driver 209 in their kernel space, the TPM also effectively protects the driver 209 and the OS 211.
[0039] Since the communication channel 215 for the DP accelerator 105-107 can be exclusively occupied by the OS 211 and the driver 209, the communication channel 215 can be protected by the TPM chip. In one embodiment, the communication channel 215 includes a peripheral component interconnect or a peripheral component express interconnect (PCIE) channel. In one embodiment, the communication channel 215 is a fuzzy communication channel. The communication channel can be connected to one or more hardware communication ports accessible by the driver 209 for communicating with the DP accelerator 105-107 through the communication channel 215. As described herein, the communication channel 215 can be protected using a session key. Each communication channel 215 can be protected using a session key different from other communication channels 215. The driver 209 may include mapping each of the DP accelerators 105-107 to a hardware communication port and an adjacency table of session keys associated with each hardware communication port.
[0040] The host machine 104 may include a trusted execution environment (TEE) 210 that is enforced to be protected by a TPM / CPU 213. A TEE is a secure environment. A TEE can ensure that the code and data loaded inside the TEE are protected in terms of confidentiality and integrity. An example of a TEE may be Software Guard Extensions (SGX) or Secure Encrypted Virtualization (SEV). SGX and / or The SEV may include a collection of central processing unit (CPU) instruction codes that allow user-level code to allocate a private area of the CPU's memory that is protected from processes running at a higher privilege level. Here, the TEE 210 may protect a user application 205 and a runtime library 206, where the user application 205 and the runtime library 206 may be provided by an end user and a DP accelerator vendor, respectively. Here, the runtime library 206 may convert application programming interface (API) calls into commands for execution, configuration, and / or control of the DP accelerator. In one embodiment, the runtime library 206 provides a set of predetermined (e.g., predefined) kernels for execution by user applications. In an embodiment, the kernel may be stored as a kernel 203 in (one or more) storage devices 204.
[0041] The host machine 104 may include a memory safety application 207 implemented in a memory safety language such as Rust and GoLang. Memory security like These memory security applications running on the release can also protect the system 200 from data confidentiality and integrity attacks. However, the operating system can be any Release, OS or OS.
[0042] The host machine 104 may be configured as follows: The distribution is installed on a system equipped with TPM secure boot. The installation can be performed offline during the manufacturing or preparation phase. The installation can also ensure that the user space applications of the host system are programmed using a memory-safe programming language. Ensuring that other applications running on the host system 104 are memory-safe applications can further mitigate potential confidentiality and integrity attacks on the host system 104.
[0043] After installation, the system can then be started via a TPM-based secure boot. TPM secure boot ensures that only signed / certified operating systems and accelerator drivers are turned on in the kernel space that provides accelerator services. In one embodiment, the operating system 211 can be loaded via a hypervisor (not shown). A hypervisor or virtual machine manager is computer software, firmware, or hardware that creates and runs virtual machines. Kernel space is a declarative area or range where a kernel (i.e., a set of predetermined (e.g., predefined) functions for execution) is identified to provide functions and services to user applications. In the event that the integrity of the system is compromised, TPM secure boot may not be able to start, but instead shut down the system.
[0044] After the secure boot, the runtime library 206 runs and creates the TEE 210, which places the runtime library 206 in the trusted memory space associated with the CPU 213. Next, the user application 205 is opened in the TEE 210. In one embodiment, the user application 205 and the runtime library 206 are statically linked and opened together. In another embodiment, the runtime library 206 is first opened in the TEE 210, and then the user application 205 is dynamically loaded in the TEE 210. In another embodiment, the user application 205 is first opened in the TEE, and then the runtime library 206 is dynamically loaded in the TEE 210. The static link library is a library that is linked to the application at compile time. Dynamic loading can be performed by a dynamic linker. The dynamic linker loads and links shared libraries for running user applications at runtime. Here, the user application 205 and the runtime library 206 in the TEE 210 are visible to each other at runtime, for example, all process data are visible to each other. However, external access to the TEE is denied.
[0045] In one embodiment, the user application 205 can only call kernels from a set of kernels predetermined by the runtime library 206. In another embodiment, the user application 205 and the runtime library 206 are hardened using a side-channel-free algorithm to resist side-channel attacks such as cache-based side-channel attacks. A side-channel attack is any attack based on information obtained from the implementation of a computer system, rather than based on weaknesses in the implemented algorithm itself (e.g., cryptanalysis and software errors). Examples of side-channel attacks include cache attacks, which are attacks based on the ability of an attacker to monitor the cache of a shared physical system in a virtualized environment or cloud environment. Hardening may include shielding the cache, outputs generated by the algorithm to be placed on the cache. Next, when the user application completes execution, the user application terminates its execution and exits from the TEE.
[0046] In one embodiment, the TEE 210 and / or the memory security application 207 are not required, for example, the user application 205 and / or the runtime library 206 are hosted in the operating system environment of the host 104.
[0047] In one embodiment, the set of kernels includes an obfuscation kernel algorithm. In one embodiment, the obfuscation kernel algorithm can be a symmetric or asymmetric algorithm. A symmetric obfuscation algorithm can use the same algorithm to obfuscate and deobfuscate data communications. An asymmetric obfuscation algorithm requires a pair of algorithms, wherein the first of the pair of algorithms is used for obfuscation and the second of the pair of algorithms is used for deobfuscation, or vice versa. In another embodiment, the asymmetric obfuscation algorithm includes a single obfuscation algorithm for obfuscating a data set, but the data set is not intended to be deobfuscated, for example, there is no corresponding deobfuscation algorithm. Obfuscation refers to obfuscating the intended meaning of communication by generally using confusing and ambiguous language to make the communication message difficult to understand. For reverse engineering, obfuscated data is more difficult and more complex. An obfuscation algorithm can be applied to obfuscate (encrypt / decrypt) data communications before conveying data, thereby reducing the chance of eavesdropping. In one embodiment, the obfuscation algorithm may also include an encryption scheme to further encrypt the obfuscated data for an additional layer of protection. Unlike encryption, which can be computationally intensive, an obfuscation algorithm can simplify calculations. Some obfuscation techniques may include, but are not limited to, letter obfuscation, name obfuscation, data obfuscation, control flow obfuscation, and the like. Letter obfuscation is a process of replacing one or more letters in data with specific replacement letters, thereby rendering the data meaningless. Examples of letter obfuscation include letter rotation functions, where each letter is shifted or rotated a predetermined number of positions along the alphabet. Another example is to reorder or scramble letters based on a specific pattern. Name obfuscation is the process of replacing a specific target string with a meaningless string. Control flow obfuscation can change the order of control flow in a program with additive code (inserting dead code, inserting uncontrolled jumps, inserting replacement structures) to hide the true control flow of the algorithm / AI model. The following article describes a system and method for sharing a key for obfuscation.
[0048] In summary, system 200 provides multiple layers of protection for the DP accelerator (for data transmission including machine learning models, training data, and inference outputs) to avoid loss of data confidentiality and integrity. System 200 may include a TPM-based secure boot protection layer, a TEE protection layer, and a kernel authentication / verification layer. In addition, system 200 may provide a memory-safe user space by ensuring that other applications on the host machine are implemented in a memory-safe programming language, which can further eliminate attacks by eliminating potential memory corruption / vulnerabilities. In addition, system 200 may include applications using side-channel-free algorithms to defend against side-channel attacks, such as cache-based side-channel attacks.
[0049] The runtime library 206 may provide an obfuscation kernel algorithm to obfuscate data communications between the host 104 and the DP accelerators 105-107. In one embodiment, obfuscation may be paired with a cryptographic scheme. In another embodiment, obfuscation is a separate protection scheme and is rendered unnecessary for the DP accelerator based on cryptographic hardware.
[0050] Figure 2B is a block diagram illustrating an example of a host channel manager (HCM) 259 communicatively coupled to one or more accelerator channel managers (ACMs) 270 connected to DP accelerators 105-107 in accordance with some embodiments. Figure 2B In one embodiment, the HCM 259 includes an authentication module 251, a termination module 252, a key manager 253, (one or more) key stores 254, and a cryptographic engine 255. The authentication module 251 can authenticate the user application running on the host server 104 to obtain permission to access or use the resources of the DP accelerator 105. The termination module 252 can terminate the connection (e.g., the channel associated with the connection will be terminated). The key manager 253 can manage (e.g., create or destroy) asymmetric key pairs or symmetric keys used to encrypt / decrypt one or more data packets of different secure data exchange channels. Here, each user application (as Figure 2A The secure data exchange channels (part of user application 205) may correspond to or be mapped to different secure data exchange channels in a one-to-many relationship, and each data exchange channel may correspond to DP accelerator 105. Each application may use multiple session keys, where each session key is used for a secure channel corresponding to a DP accelerator (e.g., accelerators 105-107). (One or more) key stores 254 may store encrypted asymmetric key pairs or symmetric keys. The cryptographic engine 255 may encrypt or decrypt data packets for data exchanged through any of the secure channels. Note that some of these modules may be integrated into fewer modules.
[0051] In one embodiment, the DP accelerator 105 includes an ACM 270 and a security unit (SU) 275. The security unit 275 may include a key manager 271, (one or more) key storages 272, a true random number generator 273, and a cryptographic engine 274. The key manager 271 may manage (e.g., generate, securely maintain and / or destroy) asymmetric key pairs or symmetric keys. (One or more) key storages 272 may store cryptographic asymmetric key pairs or symmetric keys in a secure storage within the security unit 275. The true random number generator 273 may generate seeds for key generation and use by the cryptographic engine 274. The cryptographic engine 274 may encrypt or decrypt key information or data packets for data exchange. In some embodiments, the ACM 270 and the SU 275 are integrated modules.
[0052] The DP accelerator 105 may also include memory / storage 280, which may store (one or more) artificial intelligence models 277, (one or more) watermark kernels 278 (including inherited watermark kernels, watermark-enabled kernels, watermark signature kernels, etc.), encryption and decryption kernels 281, and data 279. The HCM 259 may communicate with the ACM 270 via the communication channel 215.
[0053] Figure 3 3 is a block diagram showing a method 300 for generating a session key for protecting communication by a host and a virtual data processing accelerator (VDP accelerator or VDPA) or two virtual data processing accelerators according to an embodiment. Method 300 can be used between a first virtual data processing (VDP) accelerator "Accelerator 1" and a second node "Node 2". Node 2 can be a host device or a second VDP accelerator. Accelerator 1 has a rooted key pair PK_RK1 and SK_RK1. PK_RK1 is the public key of the rooted asymmetric key pair (RK1) of accelerator 1. SK_RK1 is the private (secret) key (SK) of the rooted asymmetric key pair (RK1) of accelerator 1. The rooted key pair RK1 is stored in the secure storage of accelerator 1. Similarly, node 2 (host or another VDP accelerator) has a rooted key pair PK_RK2 and SK_RK2. RK2 can be stored in the secure storage of node 2.
[0054] In operation 301, the accelerator 1 generates a derived asymmetric key pair PK_D1 and SK_D1 from the rooted key pair PK_RK1 and SK_RK1. Deriving an asymmetric key pair is known in the art and will not be described herein.
[0055] In operation 302, accelerator 1 sends a "get public key" command (GET_PUB_KEY) to node 2 to request the public key of node 2. GET_PUB_KEY includes two public keys of encryption accelerator 1: PK_RK1 and PK_D1. In an embodiment, PK_RK1 and PK_D1 can be encrypted using the private root key SK_RK1 of accelerator 1. The GET_PUB_KEY command also includes the public root key of accelerator 1, PK_RK1 in plain text. Node 2 can use PK_RK1 to decrypt the encrypted key of accelerator 1 and verify that the GET_PUB_KEY request is indeed from accelerator 1.
[0056] In operation 303, node 2 generates a derived asymmetric key pair PK_D2 and SK_D2 from the rooted key pair PK_RK2 and SK_RK2 of node 2. The derived keys PK_D2 and SK_D2 may be stored in a secure storage at node 2.
[0057] In operation 304, node 2 may decrypt the "GET_PUB_KEY" command received from accelerator 1 using the plaintext public root key PK_RK1 of accelerator 1. Once decrypted, node 2 obtains the derived public key of accelerator 1: PK_D1.
[0058] In operation 305, node 2 sends a "return public key" (RET_PUB_KEY) message to accelerator 1. The message includes node 2's PK_RK2 and PK_D2 encrypted using node 2's private root key SK_RK2. Node 2's public root key PK_RK2 is encapsulated with the encrypted keys PK_RK2 and PK_D2, and the encapsulated keys are then encrypted using accelerator 1's derived public key PK_D1.
[0059] In operation 306, accelerator 1 decrypts the RET_PUB_KEY message using the private derived key SK_D1 of accelerator 1. After decryption, accelerator 1 can obtain the public root key PK_RK2 of node 2. Then, accelerator 1 decrypts the encrypted keys PK_RK2 and PK_D2 using the newly obtained public root key PK_RK2 of node 2. Then, accelerator 1 can obtain the derived public key PK_D2 of node 2. In an embodiment, accelerator 1 can verify either or both of PK_RK2, decrypted PK_RK2, and plaintext PK_RK2 by checking with the host device or a historical copy of PK_RK2.
[0060] In operation 307 , accelerator 1 may generate a random number “ nc1 ”.
[0061] In operation 308, accelerator 1 may send a command "generate session key" (CMD_SESS_KEY) to node 2. The command includes a random number nc1 encrypted using the public derived key PK_D2 of node 2. CMD_SESS_KEY instructs node 2 to generate a session key from the random number nc1 of accelerator 1 and the random number nc2 generated by node 2.
[0062] In operation 309 , node 2 may decrypt the random number nc1 in the received CMD_SESS_KEY using the privately derived key SK_D2 of node 2 .
[0063] In operation 310, node 2 may generate a random number nc2. Node 2 may then generate a session key based on the random numbers nc1 and nc2. Node 2 stores the session key in an adjacency table of node 2. The session key is stored in association with accelerator 1 and a unique identifier of accelerator 1.
[0064] In operation 311, node 2 may send the random number nc2 to accelerator 1. Node 2 encapsulates nc1, nc2, and PK_D1 in a first packet, and encrypts the first packet using a privately derived key SR_D2 of node 2. Node 2 then adds PK_D2 to the encrypted first packet, and generates a second encrypted packet encrypted using a public derived key PK_D1 of accelerator 1. The encrypted second packet is then sent to accelerator 1.
[0065] In operation 312, accelerator 1 receives the encrypted second packet from node 2 and decrypts the second packet using the derived private key SK_D1 of accelerator 1. Accelerator 1 can then remove PK_D2 from the decrypted second packet, leaving only the encrypted first packet. In an embodiment, accelerator 1 can verify that PK_D2 removed from the decrypted second packet matches PK_D2 previously received in operation 305 and decrypted in the above operation 306. Accelerator 1 can also verify that nc1 obtained from the decrypted first packet and previously sent to node 2 in operation 308 has not expired (also known as "verifying freshness"). Accelerator 1 can then generate a session key based on random numbers nc1 and nc2. Accelerator 1 can store the generated session key in an adjacency table of accelerator 1 in association with the unique identifier of node 2 and the session key.
[0066] At this point, both accelerator 1 and node 2 have the same session key derived from random numbers nc1 and nc2, and both accelerator 1 and node 2 have stored the session key in their respective adjacency tables. The adjacency table will be referred to below. Figure 5 Detailed description.
[0067] Figure 44 is a block diagram showing a hardware configuration 400 of a host computing device 104 and a plurality of data processing (DP) accelerators 105-107. Each DP accelerator, such as 105, may be configured as a plurality of, for example, four virtual data processing (VDP) accelerators 105A-105D. Each VDP accelerator in the DP accelerator may communicate with another VDP accelerator in a different DP accelerator via a virtual communication channel 402. For example, VDP accelerator 105A may be configured to communicate with host 104 and VDP accelerators 106A and 107A. VDP accelerator 105B may be configured to communicate with host 104 and VDP accelerators 106B and 107B. VDP accelerator 105C may be configured to communicate with host 104 and VDP accelerators 106C and 107C. VDP accelerator 106A is also configured to communicate with VDP accelerator 107A. VDP accelerator 106B is also configured to communicate with VDP accelerator. VDP 106C is also configured to communicate with VDP 107C, and VDP 106D is configured to communicate with VDP 107D. These are merely example configurations.
[0068] To implement the configuration, each of the host 104 and the DP accelerator may generate one or more adjacency tables indicating the VDP accelerators configured within the DP accelerator. The VDP accelerators in the DP accelerator may also each have an adjacency table indicating other VDP accelerators with which each VDP accelerator may communicate, a virtual communication channel 402 used by the VDP accelerator for communication, and a session key for protecting communications on the virtual communication channel 402.
[0069] Host 104 is communicatively coupled to each of DP accelerators 105, 106, and 107. DP accelerator 105 is communicatively coupled to DP accelerators 106 and 107. Host 104 includes a communication interface having, for example, ports 0, 1, and 2. Figure 4, communication port 0 of DP accelerator 105 is communicatively coupled to communication port 0 of host 104. Communication port 0 of DP accelerator 106 is communicatively coupled to communication port 1 of host 104. Communication port 0 of DP accelerator 107 is communicatively coupled to communication port 2 of host 104. DP accelerators 105-107 are also communicatively coupled to each other. Communication port 1 of DP accelerator 105 is communicatively coupled to communication port 1 of DP accelerator 106. Communication port 2 of DP accelerator 105 is communicatively coupled to communication port 2 of DP accelerator 107. Communication port 2 of DP accelerator 106 is communicatively coupled to communication port 1 of DP accelerator 107. Each of the above physical ports is communicatively coupled to define a physical communication channel 401. Each physical communication channel 401 may be divided into a plurality of virtual communication channels 402. The number of virtual communication channels 402 per physical channel 401 may be greater than or equal to the number of VDP accelerators with which the DP accelerator is configured. For example, if the DP accelerator 105 is configured with four virtual DP accelerators, the communication ports 0, 1, and 2 of the DP accelerator 105 may each be configured with four or more virtual communication channels 402, such that there is a virtual communication channel 402 associated with each virtual VDP accelerator 105A-105D of the DP accelerator 105. One or more additional virtual communication channels 402 may be configured between the host 104 and any one of the DP accelerators 105-107, or between any one of the DP accelerators 105-107. Such additional virtual communication channels 402 may be used to send out-of-band information between the host 104 and the DP accelerator or between the DP accelerators. In an embodiment, the host 104 may communicate with the DP accelerator 105 to send configuration information to the DP accelerator 105, including instructions for the DP accelerator 105 to configure itself for the identified multiple VDP accelerators, each VDP accelerator being assigned to one or more virtual communication channels 402. Each virtual communication channel 402 may have a different session key than any other of the virtual communication channels 402. Similarly, host 104 instructs other DP accelerators (106, 107, etc.) to configure themselves and their VDP accelerators and virtual communication channels as described above for DP accelerator 105.
[0070] Each of the aforementioned virtual communication channels 402 may be protected by a different session key than that used for other virtual communication channels 402. Therefore, if any of the virtual communication channels 402 is destroyed, the other virtual communication channels 402 are still secure. In addition, there is redundancy in the communication with respect to the host device 104. Each DP accelerator 105-107 may monitor its own physical communication channel 401 to ensure that each communication channel 401 is operable. Each virtual DP accelerator may also monitor the virtual communication channel 402 to which the virtual DP accelerator is connected to ensure that each virtual communication channel 402 is operable. If a virtual communication channel 402 fails, the VDP accelerator that detects the failure may notify the DP accelerator in which the VDP accelerator is configured of the failure. The DP accelerator may then notify the host 104 of the failure using one of the additional virtual communication channels 402 reserved for out-of-band communication. Out-of-band communication is communication for purposes other than communication related to the work between the host 104 and the VDP accelerator or between two VDP accelerators.
[0071] Each of the host 104 and the virtual DP accelerators 105A-D, 106A-D, and 107A-D may have an adjacency table that stores a list of nodes (hosts or VDP accelerators) to which the host 104 or VDP accelerator 105-107 is communicatively coupled. The adjacency table will be referred to below. Figure 5 describe.
[0072] Figure 5 is a block diagram illustrating a secure communication adjacency table 500 between a host 104 and multiple virtual DP accelerators (VDP accelerators), and example adjacency tables 510A and 510B for two VDP accelerators 105A and 105B, respectively.
[0073] As above Figure 4 As shown, the host 104 and the DP accelerators 105-107 are communicatively coupled via communication ports on each of the host and the DP accelerator. Each DP accelerator can be configured to have multiple virtual data processing (VDP) accelerators. The VDP accelerator communicates with the host 104 via a virtual communication channel 402 implemented on a physical communication channel 401 between the host 104 and the communication port of the DP accelerator, and communicates with other VDP accelerators of other DP accelerators.
[0074] Host 104, for example, may have an adjacency table 500 that lists VDP accelerators that are communicatively coupled to host 104. VDP accelerators, such as 105A-105D, 106A-106D, and 107A-107D, may each have a unique virtual device ID (VID) 501, such as VDP_105A_ID, etc., so that the VDP accelerator can be referenced by name. In an embodiment, when a host wants to send a message to a VDP accelerator, the message may have the format [source, message_payload, destination]. The message may be any payload specified by the sender. An example payload includes instructions to the VDP accelerator to cause the VDP accelerator to configure itself for secure communication with another node (host or VDP accelerator). The payload may include a computational task sent from a VDP accelerator to another VDP accelerator. The payload may include another VDP accelerator sending a result back to the VDP accelerator that assigns a task to another VDP accelerator. The payload can also be out-of-band communication between the host and the VDP Accelerator or between two VDP Accelerators. Out-of-band communication can be any communication that is not related to performing end-user or application processing work.
[0075] A host as a sender (source) may refer to itself by its own ID 501, such as HOST_104_ID. A host 104 may refer to a destination VDP accelerator by its unique VID 501 or by a virtual communication channel identifier 503 of the host, which is associated with the destination VDP accelerator. The host adjacency table 500 may also include a physical port reference 502. Thus, if a host 104 with a VID 501 of HOST_104_ID sends a message to a VDP accelerator VDPA_106C, the host may look up the VID 501 of VDP accelerator 106A to determine the virtual channel 503 associated with the destination VDPA_106A (here, "1.0" or "0th virtual communication channel 402 of physical port 1 of host 104"). The message may be encrypted using a session key 504 for the host and VDP accelerator 106A (here, "0x51BF2318..." as an example). The host adjacency table 500 may also include contextual information 505 related to the host and a particular VDP accelerator that is communicatively coupled to the host via the virtual communication channel 402 .
[0076] The context information may include an identifier of a host process ("Host_Proc_ID") that has called the VDP Accelerator to perform work on behalf of the host 104. The context information may also be an identifier of a process on the VDP Accelerator that is performing work on behalf of the host 104 (e.g., "VDPA_Proc_ID"). The context information may also include a status of the host process, such as "waiting for results from the VDP Accelerator process" and / or a status of the VDP Accelerator process that is performing work on behalf of the host 104, such as "waiting for VDP Accelerator 106B subtask results" or other status. The context information may also include an identifier of a user associated with the host process (User_ID). In general, the context information may be used to answer user queries at the host, such as "what is the status of the artificial intelligence (AI) model inference I requested". Logic on the host 104 may query an adjacency table to find context information for a user with a particular User_ID, and then determine the status of each VDP Accelerator and the related process that performs AI model inference for the user. Determining such a status may include an out-of-band query of one or more VDP Accelerators to determine whether a particular subtask of a process at a particular VDP Accelerator has failed. Other context information may be generated that may be used to assist one or more VDP accelerators in diagnosing the execution of one or more processing tasks on behalf of a caller (eg, host 104) to the one or more VDP accelerators.
[0077] In an embodiment, the host delegates work to a single VDP accelerator using instructions to delegate one or more subtasks of work to one or more VDP accelerators. In an embodiment, the single VDP accelerator may determine an idle VDP accelerator that can perform the one or more subtasks, and the VDP may determine a specific subtask to delegate to the idle VDP accelerator. In an embodiment, the host 104 may provide instructions to the single VDP accelerator regarding which VDP accelerator(s) to delegate the one or more subtasks to, and the single VDP accelerator may override the host's instructions to facilitate the single VDP accelerator determining that one or more VDP accelerators are currently idle and available.
[0078] Similar to host 104, a DP accelerator, such as DP accelerator 105, may have multiple adjacency tables 510A-510D, one for each of VDP accelerators 105A-105D of DP accelerator 105. Figure 5, example VDP accelerator adjacency tables 510A and 510B are shown for VDP accelerators 105A and 105B. VDP accelerator adjacency tables for any other VDP accelerators (105C, 105D, 106A-106D, and 107A-107D) are similarly derived. The adjacency table 510A for VDP accelerator 105A may include an endpoint virtual ID 511A (e.g., host 104, or adjacent VDP accelerators 106A and 107A), a physical port 512A identifier of the port of DP accelerator 105 hosting VDP accelerator 105A, a virtual channel 513A indicating a virtual communication channel 402 used by VDP accelerator 105A to communicate with adjacent node 511A (host 104, VDPA_106A, or VDPA 107A), and a session key 514A to be used when communicating with the adjacent node having VID 511A via the virtual communication channel. The adjacency table for VDP 105B may include a physical port 512A identifier of the port of DP accelerator 105 hosting VDP accelerator 105A, a virtual channel 513A indicating a virtual communication channel 402 used by VDP accelerator 105A to communicate with adjacent node 511A (host 104, VDPA_106A, or VDPA 107A), and a session key 514A to be used when communicating with the adjacent node having VID 511A via the virtual communication channel. Figure 5 , shown as 510B, is derived similarly.
[0079] Each adjacency table (e.g., 510A) may also have context information (e.g., 505A) for processing tasks that may be executed within an adjacent VDP accelerator. Context information 505A may include, for example, an identification of a node (host or VDP accelerator) that called the VDP accelerator, such as host 104 that called VDP accelerator 105A. In table 510A, the process with the process_ID of host 104 is the caller of VDP accelerator 105A to perform work on behalf of the host process "Host_104_ID:Process_ID". The caller state of Host_104_ID:Process_ID may be that the process is waiting for a result to be returned from VDP accelerator 105A. In adjacency table 510A, VDP accelerator 106A may have been called by VDP accelerator 105A to perform work that forms a subtask of the processing task that VDP accelerator 105A is performing for the calling process Host_104_ID:Process_ID. The caller state of VDP Accelerator 105A with respect to called VDP Accelerator 106A is that VDP Accelerator 105A is waiting for a result to be returned from VDP Accelerator 106A before VDP Accelerator 105A can return the result to Host_104_ID:Process_ID. VDP Accelerator 107A adjacent to VDP Accelerator 105A may be idle and thus have no identified caller.
[0080] A session key may be generated for each virtual communication channel 402 between two nodes (host 104 to VDP accelerator or VDP accelerator to VDP accelerator). Figure 3Describes the generation of a session key between two nodes, see below Figure 6 A method is described. A session key of NULL indicates that a session key has not been determined between two nodes referenced in a row item of an adjacency table having a NULL session key. When a DP accelerator receives instructions and configuration information from a host 104 for configuring a VDP accelerator on the DP accelerator and generating an adjacency table for each VDP accelerator, the session key field for each row item is initially set to NULL. When a session key is generated for a row item in the adjacency table, the NULL is replaced with the generated session key. When all session keys have been generated for all row items of the adjacency table in all VDP accelerators of the DP accelerator, the DP accelerator may send an indication to the host 104 that the DP accelerator and its corresponding VDP accelerator are ready to receive a processing task. When all DP accelerators send such an indication to the host 104, the host 104 may perform the processing task using all configured VDP accelerators.
[0081] Figure 6 is a block diagram illustrating a method 600 of a host device instructing a plurality of virtual data processing accelerators to configure themselves for secure communications according to an embodiment.
[0082] In operation 601, a host device determines a list of data processing accelerators (DPAs) communicatively coupled to a communication port of a host. The determination may be made from, for example, an administrator's configuration file that lists the ports of the DPAs and DPA identifiers. In an embodiment, the host may send a query using each host communication port to determine one or more DPAs connected to each communication port of the host. The host may then send instructions (and optionally configuration information) instructing each DPA to configure itself with multiple virtual data processing accelerators (VPDAs) and to generate an adjacency table (e.g., VDPAs) for each of the DPAs. Figure 5 , Table 501A or 501B).
[0083] In operation 602, a host (e.g., host 104) generates and stores an adjacency table that lists each virtual data processing accelerator (VDP A) configured to communicate with the host via a virtual communication channel. In an embodiment, one or more VDP accelerators may be configured by a system administrator using a configuration file. The configuration file may indicate which VDP accelerators may communicate with which other VDP accelerators. The configuration file may specify unique identifiers for the host and the VDP accelerator, a specific communication port number to which each DP accelerator is assigned, and a virtual communication channel for a specific communication port to communicate with the VDP accelerator. There may be any number of VDP accelerators. For simplicity, a host 104 and three DP accelerators, such as 105-107, are described, where each DP accelerator is configured for four VDP accelerators. The description of three DP accelerators, each with four VDP accelerators, is exemplary and non-limiting. The system may be configured with any number of hosts and any number of DP accelerators, each DP accelerator being configured for any number of VDP accelerators. The adjacency table generated for the host may be similar to the above reference. Figure 5 Described host table 500.
[0084] In operation 603, logic in the host may iterate through a list of VDP accelerators configured to communicate with the host over a virtual communication channel. For each VDP accelerator, operations 604 through 606 may be performed. If there are no more VDP accelerators in the list, the method 600 ends.
[0085] In operation 604, the host selects a VDP accelerator from the list and generates a session key with the selected VDP accelerator. Figure 3 Described is the generation of a session key between a VDP Accelerator and a host (node). The host stores the generated session key in an entry of an adjacency table corresponding to a selected VDP Accelerator. The host uses an administrator configuration file to complete a line item of an entry in the adjacency table for the selected VDP Accelerator, including a unique identifier for the VDP Accelerator, a port number of the host for communicating with the DP Accelerator, and a virtual communication channel identifier, and any relevant context information, if any, for the selected VDP Accelerator.
[0086] In operation 605, the host sends an instruction to the selected VDP accelerator for the VDP accelerator to create its own adjacency table. The information in the instruction sent by the host may be obtained from an administrator configuration file. The instruction includes a list of other VDP accelerators that the selected VDP accelerator is to include when the selected VDP accelerator generates its own adjacency table. The instruction may also include a unique identifier for each of the other VDP accelerators, and a port number of the selected VDP accelerator assigned to each of the other VDP accelerators, a virtual communication channel identifier of a virtual communication channel that the VDP accelerator is to use to communicate with one of the other VDP accelerators. The VDP accelerator stores the generated session key for each VDP accelerator in its own row item in the adjacency table, with the row item corresponding to the VDP identifier of each VDP accelerator (e.g., Figure 5 The instructions also include instructions for the selected VDP accelerator to generate a session key with each of the other VDP accelerators and store the session key in an adjacency table of the selected VDP accelerator. Figure 7 A method is described for a selected VDP accelerator to generate its own adjacency table.
[0087] In operation 606, the host receives a signal from the selected VDP accelerator that the selected VDP accelerator has generated its own adjacency table, populated the adjacency table using the information provided in operation 605 above, and has generated and stored session keys for each of the other VDP accelerators in the adjacency table of the selected VDP accelerator. In an embodiment, the DP accelerator may monitor the completion of operations 603-605, and when operations 603-605 are completed for all VDP accelerators configured for the DP accelerator, the DP accelerator may send a single message to the host that each of the VDP accelerators of the DP accelerator is configured and has a session key. Method 600 continues at operation 603.
[0088] Figure 7 is a block diagram illustrating a method 700 of a virtual data processing accelerator configuring itself for secure communication with one or more other virtual data processing accelerators according to an embodiment.
[0089] In operation 701, the DP accelerator receives an instruction from a host device to generate a plurality of virtual DP accelerators (VDP accelerators) and generates an adjacency table of communicatively connected nodes (hosts or VDP accelerators) for each of the generated VDP accelerators. The exemplary adjacency table is described above with reference to Figure 5 Description. The information in the instruction sent by the host can be obtained by the host from a configuration file created by an administrator. In an embodiment, the instruction can be a default instruction.
[0090] In operation 702, for each of the VDP accelerators generated by the DP accelerator, a VDP accelerator is selected, and the DP accelerator logic performs or causes the selected VDP accelerator to perform the following operations 703 through 706. When there are no more VDP accelerators to process, the method 700 ends.
[0091] In operation 703, the selected VDP accelerator receives an instruction to generate an adjacency list for the selected VDP accelerator, the adjacency list listing a plurality of nodes (hosts or other VDP accelerators) to which the selected VDP is communicatively coupled via a virtual communication channel between the selected VDP accelerator and one of the plurality of nodes.
[0092] In operation 704, the VDP accelerator generates and stores an adjacency table for the selected VDP accelerator according to the instruction received in operation 703. Figure 5 In the example of FIG. 1 , the adjacency table may include a row entry for each of the other nodes (hosts or VDP accelerators) to which the selected VDP accelerator is communicatively coupled. Information for each row entry may include a virtual device identifier (VID, e.g., Figure 5 , table 510A, VID 511A), and the physical port of the DP accelerator hosting the selected VDP accelerator (e.g. Figure 5 512A) corresponds to the selected VDP accelerator's communication port, virtual communication channel identifier (e.g. Figure 5 513A), and a session key field for storing a session key generated for communication between the selected VDP accelerator and one of the other nodes.
[0093] In operation 705, the selected VDP Accelerator generates a session key for each of the other nodes (hosts or VDP Accelerators) configured to communicate with the selected VDP Accelerator. The session key is stored in a row item entry of an adjacency table of the selected VDP Accelerator, the row item entry being associated with the node. The selected VDP Accelerator generates a different session key between the selected VDP Accelerator and each of the other neighboring nodes (hosts or VDP Accelerators) to the selected VDP Accelerator. The session key is for use in communication between the selected VDP Accelerator and one of the other nodes. Figure 3 Generating a session key between a VDP Accelerator and a node (host or VDP Accelerator) is described. The VDP Accelerator logic stores the session key in association with the selected VDP Accelerator in its adjacency table for the selected VDP Accelerator.
[0094] In operation 706, if there are no more VDP accelerators to be configured for the DP accelerator, and therefore no more session keys to be generated for the VDP accelerator of the DP accelerator, the DP accelerator sends a message or signal to the host that the DP accelerator has completed configuring the VDP accelerator and generating an adjacency table for the VDP accelerator and a session key for secure communication with each other VDP accelerator in the adjacency table. In an embodiment, each session key in the adjacency table is different from the other session keys in the adjacency table. In an embodiment, each configured VDP accelerator of the DP accelerator separately notifies the host 104 that the VDP accelerator has been configured and the session key generated for the VDP accelerator. In an embodiment, each configured VDP accelerator notifies the DP accelerator hosting the VDP accelerator that the VDP accelerator is configured and that a session key has been generated for the VDP accelerator. Then, when all VDP accelerators of the selected DP accelerator are configured and the session key is generated, the DP accelerator notifies the host 104. The method 700 continues at operation 702.
[0095] Figure 8 is a block diagram illustrating a method 800 of a virtual data processing (VDP) accelerator receiving a processing task from a host and executing one or more subtasks of the task through one or more additional VDP accelerators according to an embodiment.
[0096] In operation 801, the VDP Accelerator receives a processing task from a host device. In an embodiment, the processing task includes instructions for dividing the processing task into subtasks to be processed on at least one additional VDP Accelerator, and the VDP Accelerator has an entry in an adjacency table of the VDP Accelerator for securely communicating with the at least one additional VDP Accelerator. In this embodiment, it is assumed that the host determines that the at least one additional VDP Accelerator is idle or will be idle soon, so that the at least one additional VDP Accelerator can perform one or more subtasks on behalf of the VDP Accelerator.
[0097] In operation 802, the VDP Accelerator sends one or more subtasks and instructions to execute the subtask(s) to at least one additional VDP Accelerator. The message to send the one or more subtasks is protected using a session key generated for use between the VDP Accelerator and the at least one additional VDP Accelerator. The at least one additional VDP Accelerator executes the one or more subtasks.
[0098] In operation 803 , the VDP accelerator also executes one or more subtasks of the received processing task.
[0099] In operation 804, the VDP Accelerator receives one or more results from the at least one additional VDP Accelerator. The VDP Accelerator completes one or more subtasks of its processing task and returns one or more results from the one or more subtasks performed by the VDP Accelerator and the one or more subtasks performed by the at least one additional VDP Accelerator to the host. Method 800 ends.
[0100] Point-to-point communication between DPAs can be extended to broadcast messages from the host to a selected number of DPAs (or a subset of DPAs or a specified DPA) using a broadcast protocol through a communication chain between DPAs or via a communication switch (such as a PCIe switch) coupled to the DPAs. Broadcasting can be used in situations where the host needs to communicate messages to multiple DPAs. Broadcast communication occurs when the host's application schedules a job to be processed by a subset of DPAs (or a specified DPA) and the application needs to send the same information (e.g., input data or model) to the subset of DPAs. The broadcast protocol needs to be secure (i.e., only a subset of DPAs can read the broadcast message) but robust (e.g., with a session-based encryption scheme).
[0101] In addition, when multiple applications / users request a broadcast, the host server can cause the DPA to generate one or more virtual DPAs (VDPAs) for each corresponding DPA, each VDPAs using a portion of the DPA's resources to serve one or more applications. The virtualization of the DPA can be managed by the virtualization module of the DPA (e.g., virtualization module 282 of FIG. 2) to provide a sandbox environment for one or more applications. In one embodiment, the host server maintains a mapping table of VDPAs mapped to one or more applications, so that the host server routes communication data for the corresponding application to the mapped VDPAs. When the application terminates, the VDPAs can be released. Adjacent VDPAs can communicate with each other as long as they are generated for the same one or more applications.
[0102] For example, refer to Figure 4, a first application of host 104 may request to broadcast to DPAs 105-107, requesting to cause host 104 to generate a first set of virtual DPAs (105A, 106A, and 107A) with virtual communication channel 402A, the virtual communication channel having the same port and configuration as physical communication channel 401 for DPAs 105-107. A second application of host 104 may request to broadcast to DPAs 105-107, requesting to cause host 104 to generate a second set of virtual DPAs (105B, 106B, and 107B). A third application of host 104 may request to broadcast to DPAs 105-107, requesting to cause host 104 to generate a third set of virtual DPAs (105C, 106C, and 107C). A fourth application of host 104 may request to broadcast to DPAs 105-107, requesting that host 104 create a fourth set of virtual DPAs (105D, 106D, and 107D), etc. Each set has a corresponding virtual communication channel 402B-D having the same ports and configuration as physical communication channel 401 for the set of VDPAs.
[0103] Fig. 9 1 is a flow chart illustrating a broadcast protocol for a host and one or more virtual data processing accelerators according to an embodiment. Broadcast communication refers to the communication of data from a host or DPA or VDP A to many listeners (e.g., many DPAs or VDPAs) (e.g., a one-to-many relationship), rather than a one-to-one relationship. A receiving VDP A may include many VDPAs, including VDPAs communicating with one or more hosts or a subset of VDPAs associated with one host communicating with a requesting application requesting the broadcast. Reference Fig. 9 , the method 900 may be performed by a host 104 and one or more VDPAs 105-107 communicatively coupled to the host 104, wherein the VDPAs are Figure 4 are shown coupled to each other in a chain configuration.
[0104] In one embodiment, at block 901, the host 104 receives a broadcast request from an application of the host 104 to broadcast to a subset of the DPAs 105-107. The broadcast request may be a request to broadcast the same or repeatable message to a subset of the DPAs communicating with the host 104. The broadcast request may identify the subset of the DPAs by a DPA identifier that is part of a list of DPAs coupled to the host 104. At block 902, the host 104 generates a list of VDPAs based on the DPAs coupled to the host 104, and generates a mapping table that maps the application to the list of generated VDPAs. The host 104 may send a request to the DPA to cause the DPA to generate a VPDA based on the mapping table. In one embodiment, the mapping table specifies a virtual communication channel (e.g., channel 402) and a port associated with the VDPA, where the VDPA uses a portion of the data processing resources of the corresponding DPA. Basically, the VDPA will work like a DPA, but provide data processing resources to many user applications or environments (each with their own virtualized DPA and set of communication channels) in a virtual sandbox.
[0105] Once the VPA is generated (e.g., in this example, 105A, 106A, and 107A are generated for the application), at block 904, the host 104 sends a request for a public key associated with the VDPAs to a subset of the VDPAs. The public key may be part of a secure key pair derived from a root key of a secure unit for each of the VDPAs. At block 905, the VDPAs generate a derived secure key pair, wherein a private key of the derived key pair is maintained locally at the VDPAs, and a public key of the derived key pair is sent to the host 104 at block 906.
[0106] At block 907, the host 104 selects a broadcast VDPA. In one embodiment, the broadcast VDPA may be selected based on the average of the shortest distances from the broadcast VDPA to each of the subset of VDPAs to minimize broadcast delay. In another embodiment, the broadcast VDPA may be selected based on the currently scheduled computational load, or the available computational capacity of the VDPA compared to the remaining VDPAs. In another embodiment, the VDPA may be randomly selected based on a random number generator. For this example, for purposes of illustration, the selected broadcast VDPA is VDPA 105A. At block 908, the host 104 sends a public key to the broadcast VDPA 105A, and optionally, generates and sends a broadcast session key to the broadcast VDPA 105A, wherein the broadcast session key is a symmetric key used to encrypt and decrypt a broadcast communication session. In another embodiment, the broadcast VDPA 105A generates the broadcast session key locally, and the broadcast VDPA 105A sends the broadcast session key to the host 104. In one embodiment, the broadcast session key may be a randomly generated session key.
[0107] At block 909, upon receiving the public keys (and optionally, the broadcast session keys), the VDPA 105A encrypts the broadcast session keys using each of the public keys to generate a set of messages for dissemination. In another embodiment, block 909 may be performed by the host 104, e.g., the host 104 encrypts the broadcast session keys using each of the public keys to generate a set of messages, and the host 104 sends the set of messages to the VDPA 105A for dissemination.
[0108] At block 910, the broadcast VDPA 105A communicates with the broadcasting VDPA 105A via a virtual channel (here, such as Figure 4 As shown, VDPA 106A is coupled at port 1 of VDPA 105A, and VDPA 107A is coupled at port 2 of VDPA 105A) broadcasts the encrypted broadcast session key to the VDPA. The broadcast VDPA 105A sends the encrypted broadcast session key to VDPA 106A at port 1 and VDPA 107A at port 2. In one embodiment, the broadcast message includes a message header, a propagation instruction for specifying that the VDPA receiving the broadcast message propagate or repeat the broadcast to the VDPAs coupled to their corresponding ports, and an instruction for terminating the propagation when any VDPA receives the broadcast message twice. Reference Figure 4 For example, when VDPA 106A receives a broadcast at port 1, VDPA 106A sends or propagates a first message to VDPA coupled at port 2 of VDPA 106A, for example, sending a broadcast message to VDPA 107A via port 1 of VDPA 107A. Because VDPA 107A receives the message twice (from VDPA 105A and VDPA 106A), the first message propagation terminates at VDPA 107A. Similarly, when VDPA 107A receives a broadcast at port 2 (because VDPA 105A sends a broadcast to VDPA 107A at port 2), VDPA 107A sends or propagates a second propagation message to VDPA coupled at port 1 of VDPA 107A, for example, sending a broadcast message to VDPA 106A via port 2 of VDPA 106A. Here, because VDPA 106A receives the message twice (from VDPA 105A and VDPA 107A), the second message propagation terminates at VDPA 106A. At block 911, upon distributing the propagation message, VDPA 105A notifies host 104 that the broadcast session key is delivered to the VDPA. At block 912, each of the subset of VDPAs that receive the encrypted broadcast session key decrypts using the VDPA's private key and obtains the broadcast session key.
[0109] Thereafter, the host 104 may encrypt the data to be broadcast by the requested application based on the broadcast session key, and may send the data from the host 104 to the broadcast VDPA 105A for dissemination, or the broadcast VDPA 105A may encrypt the data based on the broadcast session key for dissemination. Upon receiving the broadcast data, the subset of VDPAs having the broadcast session key may decrypt and obtain the broadcast data.
[0110] In another embodiment, VDPA 105A broadcasts an encrypted broadcast session key to a VDPA coupled to only one of its ports, the broadcast including instructions for the VDPA receiving the broadcast message to propagate the broadcast to the VDPA coupled to its corresponding port, and instructions for terminating the propagation when the broadcasting VDPA 105A receives the original broadcast message. In this case, the broadcast message will be traversed in a circular manner. For example, the broadcast can be traversed in a clockwise direction, such as VDPA 105A->VDPA107A->VDPA106A->VDPA 105A, in which case VDPA105A receives the original broadcast and terminates the propagation. The broadcast can be traversed in a counterclockwise direction, such as VDPA 105A->VDPA106A->VDPA 107A->VDPA105A, in which case VDPA105A receives the original broadcast and terminates the propagation. When each of the VDPAs receives the broadcast, the subset of VDPAs participating in the broadcast can decrypt based on the private key associated with the VDPAs and obtain the broadcast session key. Thereafter, any of the subset of VDPAs receiving the broadcast data can decrypt the broadcast using the broadcast session key. In some embodiments, the broadcast data can be decrypted based on the private key associated with the VDPAs. Figure 5 The adjacency table further encrypts any communication (or broadcast) between adjacent VADPs discussed above using the adjacent session keys.
[0111] Fig.10 1 is a flow chart illustrating an example method for a host to perform a broadcast according to an embodiment. Process 1000 may be performed by processing logic that may include software, hardware, or a combination thereof. For example, process 1000 may be performed by a host system, such as Figure 4 or Fig. 9 Host 104. Reference Fig.10At box 1001, processing logic determines a list of data processing accelerators (DPAs) communicatively coupled to a host and notifies an application of available DPAs. At box 1002, in response to an application requesting a broadcast to a subset of the DPA, processing logic initiates generation of a VDPA for the DPA. At box 1003, processing logic notifies the VDPA and collects public keys for a subset of the VDPA. At box 1004, processing logic sends the public keys to the broadcast VPDA. Here, the broadcast VPDA may be selected based on an average of the closest distances of the broadcast VDPA to the remaining VDPAs, or based on the availability of computing power of the DPA, or based on random selection, etc. At box 1005, processing logic receives a signal indicating that the broadcast VDPA has encrypted a broadcast session key using a public key for a subset of the VDPA to generate a set of encrypted broadcast session keys and broadcasts the set of encrypted broadcast session keys to each of one or more VDPAs via message propagation, where only a subset of the VDPAs can decrypt and obtain the broadcast session key. In another embodiment, the host 104 encrypts the broadcast session key using a public key for a subset of the VDPA and sends the encrypted key to the broadcast VDPA for dissemination.
[0112] Fig.11 1 is a flow chart illustrating an example method for a virtual data processing accelerator to perform broadcasting according to an embodiment. Process 1100 may be performed by processing logic, which may include software, hardware, or a combination thereof. For example, process 1100 may be performed by a VDPA, such as Figure 4 or Fig. 9 VDPA 105A. Reference Fig.11 At box 1101, in response to receiving a broadcast instruction from an application, processing logic encrypts a broadcast message based on a broadcast session key for a broadcast communication session, the broadcast instruction designating one or more virtual DP accelerators among a plurality of virtual DP accelerators to receive the broadcast message. At box 1102, processing logic determines one or more public keys of one or more security key pairs each associated with one of the designated virtual DP accelerators. At box 1103, processing logic encrypts the broadcast session key based on the determined one or more public keys. At box 1104, processing logic broadcasts the encrypted broadcast message and the one or more encrypted broadcast session keys to adjacent virtual DP accelerators for dissemination, wherein the designated virtual DP accelerator decrypts the encrypted broadcast session key based on a corresponding private key associated with the designated virtual DP accelerator, wherein the message is decrypted based on the broadcast session key.
[0113] In one embodiment, a virtual DP accelerator that receives a broadcast instruction from an application is designated as a broadcast virtual DP accelerator to broadcast a message. In one embodiment, an adjacent virtual DP accelerator propagates the broadcast by receiving the broadcast and forwarding the broadcast to its adjacent virtual DP accelerators.
[0114] In one embodiment, if the designated virtual DP accelerator receives the propagation broadcast, the broadcast propagation is terminated. In one embodiment, the message for the propagation broadcast includes a message source, a message context, but does not include a message destination.
[0115] In one embodiment, a non-designated virtual DP accelerator among the plurality of virtual DP accelerators receives the broadcast but does not have a corresponding key to decrypt the encrypted broadcast for generating a broadcast session key, and the non-designated virtual DP accelerator does not have access to the broadcast session key to decrypt the broadcast message for the communication session. In one embodiment, the broadcast message is communicated via a virtual communication channel associated with the virtual DP accelerator.
[0116] With respect to any of the above aspects, the host processor may be a central processing unit (CPU), and the DP accelerator may be a general purpose processing unit (GPU) coupled to the CPU via a bus or interconnect. The DP accelerator may be implemented in the form of an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) device, or other form of integrated circuit (IC). Alternatively, the host processor may be part of a main data processing system, and the DP accelerator may be one of many distributed systems as a secondary system, the main system being able to remotely offload its data processing tasks over a network (e.g., a cloud computing system such as a software as a service, or SaaS system, or a platform as a service, or PaaS system). The link between the host processor and the DP accelerator may be a peripheral component interconnect express (PCIe) link or a network connection such as an Ethernet connection.
[0117] Note that some or all of the components shown and described above may be implemented in software, hardware, or a combination thereof. For example, these components may be implemented as software installed and stored in a permanent storage device, which may be loaded and executed in a memory by a processor (not shown) to perform the processes or operations described throughout the present application. Alternatively, these components may be implemented as executable code programmed or embedded into dedicated hardware, such as an integrated circuit (e.g., a dedicated IC or ASIC), a digital signal processor (DSP), or a field programmable gate array (FPGA), which may be accessed from an application via a corresponding driver and / or operating system. In addition, such components may be implemented as specific hardware logic in a processor or processor core as part of an instruction set accessible by a software component via one or more specific instructions.
[0118] Fig.12 1 is a block diagram illustrating an example of a data processing system that may be used with one embodiment of the present invention. For example, system 1500 may represent any of the above-described data processing systems, such as, for example, the above-described client device or a server, such as, for example, the above-described host 104 or DPAs 105-107 or VDPAs 105A-107D, that perform any of the above-described processes or methods.
[0119] System 1500 may include many different components. These components may be implemented as integrated circuits (ICs), portions thereof, discrete electronic devices, or other modules suitable for circuit boards such as a motherboard or add-in card of a computer system, or as components otherwise incorporated into a chassis of a computer system.
[0120] It should also be noted that system 1500 is intended to illustrate the high-level view of many components of a computer system. However, it should be understood that in some embodiments, there may be additional components, and in addition, in other embodiments, the different arrangements of the components shown may occur. System 1500 may represent a desktop, notebook computer, tablet computer, server, mobile phone, media player, personal digital assistant (PDA), smart watch, personal communicator, gaming device, network router or hub, wireless access point (AP) or repeater, set-top box or its combination. In addition, although only a single machine or system is shown, the term "machine" or "system" should also be considered to include executing a set of (or multiple sets of) instructions individually or jointly to perform any one or more of the machines or systems in the method discussed herein.
[0121] In one embodiment, system 1500 includes a processor 1501, a memory 1503, and devices 1505-1508 connected via a bus or interconnect 1510. Processor 1501 may represent a single processor or multiple processors including a single processor core or multiple processor cores. Processor 1501 may represent one or more general-purpose processors, such as a microprocessor, a central processing unit (CPU), etc. More particularly, processor 1501 may be a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, or a processor implementing other instruction sets, or a processor implementing a combination of instruction sets. Processor 1501 may also be one or more special-purpose processors, such as an application-specific integrated circuit (ASIC), a cellular or baseband processor, a field programmable gate array (FPGA), a digital signal processor (DSP), a network processor, a graphics processor, a communication processor, a cryptographic processor, a coprocessor, an embedded processor, or any other type of logic capable of processing instructions.
[0122] Processor 1501 may be a low-power multi-core processor socket, such as an ultra-low voltage processor, and processor 1501 may act as a main processing unit and a central hub for communicating with various components of the system. Such a processor may be implemented as a system on a chip (SoC). Processor 1501 is configured to execute instructions to perform the operations and steps discussed herein. System 1500 may further include a graphics interface that communicates with an optional graphics subsystem 1504, which may include a display controller, a graphics processor, and / or a display device.
[0123] The processor 1501 may communicate with the memory 1503, which, in one embodiment, may be implemented via multiple memory devices to provide a fixed amount of system memory. The memory 1503 may include one or more volatile storage (or memory) devices, such as random access memory (RAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), static RAM (SRAM), or other types of storage devices. The memory 1503 may store information including sequences of instructions to be executed by the processor 1501 or any other device. For example, executable code and / or data of various operating systems, device drivers, firmware (e.g., an input-output basic system or BIOS), and / or applications may be loaded into the memory 1503 and executed by the processor 1501. The operating system may be any type of operating system such as, for example, from of Operating system, Mac from Apple From of Or other real-time or embedded operating systems, such as VxWorks.
[0124] System 1500 may also include IO devices, such as devices 1505-1508, including (one or more) network interface devices 1505, (one or more) optional input devices 1506, and (one or more) other optional IO devices 1507. Network interface device 1505 may include a wireless transceiver and / or a network interface card (NIC). The wireless transceiver may be a WiFi transceiver, an infrared transceiver, a Bluetooth transceiver, a WiMax transceiver, a wireless cellular telephone transceiver, a satellite transceiver (e.g., a global positioning system (GPS) transceiver) or other radio frequency (RF) transceiver, or a combination thereof. The NIC may be an Ethernet card.
[0125] The input device(s) 1506 may include a mouse, a touchpad, a touch-sensitive screen (which may be integrated with the display device 1504), a pointer device such as a stylus, and / or a keyboard (e.g., a physical keyboard or a virtual keyboard displayed as part of the touch-sensitive screen). For example, the input device 1506 may include a touch screen controller coupled to a touch screen. The touch screen and touch screen controller may detect contact and movement or interruption thereof, for example, using any of a variety of touch sensitivity technologies, including but not limited to capacitive, resistive, infrared, and surface acoustic wave technologies, as well as other proximity sensor arrays or other elements for determining one or more contact points with the touch screen.
[0126] IO devices 1507 may include audio devices. Audio devices may include speakers and / or microphones to facilitate voice-enabled functions, such as voice recognition, voice replication, digital recording, and / or telephone functions. Other IO devices 1507 may also include universal serial bus (USB) ports, parallel ports, serial ports, printers, network interfaces, bus bridges (e.g., PCI-PCI bridges), sensors (e.g., motion sensors such as accelerometers, gyroscopes, magnetometers, light sensors, compasses, proximity sensors, etc.), or combinations thereof. Device 1507 may also include an imaging processing subsystem (e.g., a camera), which may include an optical sensor, such as a charge coupled device (CCD) or a complementary metal oxide semiconductor (CMOS) optical sensor, for facilitating camera functions, such as recording photos and video clips. Certain sensors may be coupled to interconnect 1510 via a sensor hub (not shown), while other devices such as keyboards or thermal sensors may be controlled by an embedded controller (not shown), depending on the specific configuration or design of system 1500.
[0127] In order to provide persistent storage of information such as data, applications, one or more operating systems, etc., mass storage (not shown) may also be coupled to the processor 1501. In various embodiments, to enable thinner and lighter system designs and to improve system responsiveness, the mass storage may be implemented via a solid-state device (SSD). However, in other embodiments, the mass storage may be implemented primarily using a hard disk drive (HDD) with a smaller amount of SSD storage to act as an SSD cache to enable non-volatile storage of context states and other such information during power-off events so that fast power-up can occur upon restart of system activity. In addition, a flash memory device may be coupled to the processor 1501, for example, via a serial peripheral interface (SPI). The flash memory device may provide non-volatile storage of system software, including basic input / output software (BIOS) and other firmware for the system.
[0128] The storage device 1508 may include a computer-accessible storage medium 1509 (also referred to as a machine-readable storage medium or a computer-readable medium) on which is stored one or more sets of instructions or software (e.g., modules, units, and / or logic 1528) embodying any one or more of the methods or functions described herein. The processing module / unit / logic 1528 may represent any of the above-mentioned components, such as, for example Figure 4 The processing module / unit / logic 1528 may also reside completely or at least partially within the memory 1503 and / or within the processor 1501 during execution of the processing module / unit / logic 1528 by the data processing system 1500, and the memory 1503 and the processor 1501 also constitute machine-accessible storage media. The processing module / unit / logic 1528 may also be sent or received over a network via the network interface device 1505.
[0129] Computer-readable storage medium 1509 may also be used to persistently store some of the above-mentioned software functions. Although computer-readable storage medium 1509 is shown as a single medium in the exemplary embodiment, the term "computer-readable storage medium" should be considered to include a single medium or multiple media (e.g., a centralized or distributed database, and / or associated caches and servers) that store one or more sets of instructions. The term "computer-readable storage medium" should also be considered to include any medium that can store or encode an instruction set for execution by a machine and cause the machine to perform any one or more of the methods of the present disclosure. The term "computer-readable storage medium" should therefore be considered to include, but is not limited to, solid-state memory, optical and magnetic media, or any other non-temporary machine-readable medium.
[0130] The processing modules / units / logic 1528, components, and other features described herein may be implemented as discrete hardware components or integrated into the functionality of hardware components such as ASICs, FPGAs, DSPs, or similar devices. Additionally, the processing modules / units / logic 1528 may be implemented as firmware or functional circuitry within a hardware device. Furthermore, the processing modules / units / logic 1528 may be implemented in any combination of hardware devices and software components.
[0131] Note that although system 1500 is illustrated using various components of a data processing system, it is not intended to represent any particular architecture or manner of interconnecting the components; as these details are not germane to the embodiments of the present disclosure. It will also be understood that network computers, handheld computers, mobile phones, servers, and / or other data processing systems with fewer or perhaps more components may also be used with embodiments of the present disclosure.
[0132] Some portions of the foregoing detailed description have been presented in terms of algorithms and symbolic representations of operations on data bits within a computer memory. These algorithmic descriptions and representations are the means used by those skilled in the data processing arts to most effectively convey the substance of their work to others skilled in the art. An algorithm is here, and generally, conceived to be a self-consistent sequence of operations leading to a desired result. The operations are those requiring physical manipulations of physical quantities.
[0133] It should be remembered, however, that all of these and similar terms are associated with the appropriate physical quantities and are merely convenient labels applied to these quantities. Unless otherwise stated, as will be apparent from the above discussion, it should be understood that throughout the specification, discussions using terms such as those set forth in the appended claims refer to the actions and processes of a computer system or similar electronic computing device that manipulates and transforms data represented as physical (electronic) quantities within the computer system's registers and memories into other data similarly represented as physical quantities within the computer system's memories or registers or other such information storage, transmission or display devices.
[0134] The techniques shown in the figure can be implemented using code and data stored and executed on one or more electronic devices. Such electronic devices use computer-readable media, such as non-transitory computer-readable storage media (e.g., magnetic disks, optical disks, random access memory, read-only memory, flash memory devices, phase change memory) and temporary computer-readable transmission media (e.g., electrical, optical, acoustic or other forms of propagation signals - such as carrier waves, infrared signals, digital signals) to store and communicate (internally and / or through a network with other electronic devices) code and data.
[0135] The processes or methods described in the foregoing figures may be performed by processing logic including hardware (e.g., circuits, dedicated logic, etc.), software (e.g., embodied on a non-transitory computer-readable medium), or a combination of both. Although the processes or methods are described above with respect to some sequential operations, it should be understood that some of the operations described may be performed in a different order. In addition, some operations may be performed in parallel rather than sequentially.
[0136] In the foregoing description, embodiments of the present disclosure have been described with reference to specific exemplary embodiments of the present disclosure. Obviously, various modifications may be made thereto without departing from the broader spirit and scope of the present disclosure as set forth in the appended claims. Therefore, the description and the accompanying drawings should be considered illustrative rather than restrictive.
Claims
1. A computer-implemented method for broadcasting a message to one or more virtual data processing (DP) accelerators, the method comprising: encrypting the broadcast message based on a broadcast session key for the broadcast communication session in response to receiving a broadcast instruction from the application, the broadcast instruction specifying one or more virtual DP accelerators of the plurality of virtual DP accelerators to receive the broadcast message; determining one or more public keys of one or more security key pairs, each of the security key pairs being associated with one of the specified virtual DP accelerators; encrypting the broadcast session key based on the determined one or more public keys; as well as Broadcasting an encrypted broadcast message and one or more encrypted broadcast session keys to neighboring virtual DP accelerators for dissemination, wherein a designated virtual DP accelerator decrypts the encrypted broadcast session key based on a corresponding private key associated with the designated virtual DP accelerator, wherein the message is decrypted based on the broadcast session key, and the broadcast message is communicated via a virtual communication channel associated with the virtual DP accelerator based on an adjacency table including the broadcast session key. 2 . The method of claim 1 , wherein the virtual DP accelerator receiving the broadcast message is designated as a broadcast virtual DP accelerator to broadcast the message.
3. The method of claim 1, wherein the neighboring virtual DP accelerators propagate the broadcast by receiving the broadcast and forwarding the broadcast to their neighboring virtual DP accelerators.
4. The method of claim 1, wherein if the designated virtual DP accelerator receives the propagation broadcast, the broadcast propagation is terminated.
5. The method of claim 1, wherein the broadcast message for propagation includes a message source, a message context, but does not include a message destination.
6. The method of claim 1 , wherein a non-designated virtual DP accelerator of the plurality of virtual DP accelerators receives a broadcast but does not have a corresponding key to decrypt the encrypted broadcast message for generating the broadcast session key, and the non-designated virtual DP accelerator cannot access the broadcast session key to decrypt the broadcast message for a communication session.
7. An apparatus for broadcasting a message to one or more virtual data processing (DP) accelerators, comprising: a first encryption module, configured to encrypt a broadcast message based on a broadcast session key for a broadcast communication session in response to receiving a broadcast instruction from an application, the broadcast instruction specifying one or more virtual DP accelerators of the plurality of virtual DP accelerators to receive the broadcast message; a determination module for determining one or more public keys of one or more security key pairs, each of the security key pairs being associated with one of the specified virtual DP accelerators; a second encryption module, configured to encrypt the broadcast session key based on the determined one or more public keys; as well as A broadcast module for broadcasting an encrypted broadcast message and one or more encrypted broadcast session keys to neighboring virtual DP accelerators for dissemination, wherein a designated virtual DP accelerator decrypts the encrypted broadcast session key based on a corresponding private key associated with the designated virtual DP accelerator, wherein the message is decrypted based on the broadcast session key, and the broadcast message is communicated via a virtual communication channel associated with the virtual DP accelerator based on an adjacency table including the broadcast session key.
8. A non-transitory machine-readable medium having instructions stored therein, which when executed by a processor cause the processor to perform the method according to any one of claims 1-6.
9. A data processing system comprising: processor; as well as A memory is coupled to the processor to store instructions, wherein the instructions, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 6.
