Network Interface Offloading Cross-Channel Collective Operations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High Performance Computing (HPC) environments face performance degradation due to the need for fast execution of collective operations, which is hindered by Operating System noise and the processing burden on host processors, especially in large systems.
Innovation Solution
A Network Interface (NI) with multiple work queues and control circuitry that offloads cross-channel work requests by controlling the advance of work queues based on completion status, enabling efficient execution of collective operations without software intervention, utilizing Infiniband networks and incorporating computation units for calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If collective operations are executed on host processors, then computation capability is utilized, but performance degrades due to OS noise and processing burden
Solution Approach 1:
The patent extracts the execution of collective operations from the host processor and transfers it to the network interface device. The host processor generates work requests and posts them to work queues in the network interface device, but the actual execution of collective operations (such as reduction, scatter, gather) is performed by the network interface device's control circuitry and computation units, thereby removing the processing burden from the host processor and eliminating OS noise interference.
Solution Approach 2:
The network interface device is equipped with computation units that can independently execute calculations specified in work requests without requiring host processor intervention. The device manages its own work queues, controls the advancement of operations based on completion status, and handles cross-channel communication autonomously, enabling the system to serve itself without continuous host processor involvement.
2Productivity
If software intervention is used for collective operations, then flexibility is maintained, but execution speed is reduced
Solution Approach 1:
The patent replaces the software-based execution mechanism on the host processor with a hardware-based execution mechanism in the network interface device. The control circuitry and computation units in the network interface device execute collective operations directly in hardware, substituting the software intervention that previously occurred on the host processor, thereby significantly increasing execution speed while maintaining operational flexibility through programmable work requests.
3Reliability
If host processors handle all processing, then system simplicity is maintained, but performance and jitter are worsened
Solution Approach 1:
The patent segments the processing responsibilities between the host processor and the network interface device. The host processor handles high-level application logic and generates work requests, while the network interface device handles the execution of collective operations and manages data transmission. This segmentation allows each component to specialize in its designated tasks, improving overall system performance and reducing jitter by preventing the host processor from being overwhelmed by collective operation processing.
Data Source
AI summary
A Network Interface (NI) includes a host interface, which is configured to receive from a host processor of a node one or more cross-channel work requests that are derived from an operation to be executed by the node. The NI includes a plurality of work queues for carrying out transport channels to one or more peer nodes over a network. The NI further includes control circuitry, which is configured to accept the cross-channel work requests via the host interface, and to execute the cross-channel work requests using the work queues by controlling an advance of at least a given work queue according to an advancing condition, which depends on a completion status of one or more other work queues, so as to carry out the operation.