Shared File Checkpointing via Local Burst Buffer Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current high-performance computing systems face challenges in accelerating shared-file checkpointing due to the limitations of burst buffers in handling node-local checkpoint files, which are not conducive to accelerating shared-file checkpointing where all application processes write into a single file shared among all computing nodes.
Innovation Solution
A system that utilizes local burst memory storage devices on each compute node to store segments of application data and build an associated index for the shared file, allowing for asynchronous transfer of these segments and the index to a parallel file system, enabling efficient creation and restoration of shared files across multiple compute nodes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If burst buffers are used to accelerate checkpoint creation, then I/O operations are sped up, but the limited capacity of burst buffers prevents them from replacing traditional parallel file systems
Solution Approach 1:
The patent segments the checkpoint file into multiple parts distributed across different burst buffers on various compute nodes. Each burst buffer stores a segment of the shared checkpoint file, allowing the system to leverage the high-speed I/O capability of burst buffers while distributing the storage requirement across multiple nodes, thus resolving the contradiction between speed and capacity.
Solution Approach 2:
The patent introduces a distributed dimension by spreading the checkpoint file across multiple compute nodes with burst buffers. Instead of a single centralized storage system, the checkpoint is distributed across the network, utilizing both the high-speed I/O of burst buffers and the aggregated capacity of distributed storage to achieve both speed and capacity requirements.
2Ease of operation
If all processes write into a single shared file for checkpointing, then checkpoint restoration is simplified, but the writing process becomes a bottleneck that reduces computing efficiency
Solution Approach 1:
The patent segments the shared checkpoint file into multiple distributed segments stored on different compute nodes. Each compute node writes its local segment independently to its local burst buffer, eliminating the bottleneck of sequential writes to a single shared file. The segments are later assembled during restoration, maintaining simplicity while dramatically improving productivity.
Solution Approach 2:
The patent performs preliminary actions by having each compute node write its checkpoint segment locally to its burst buffer before the actual shared file is fully assembled. This allows compute nodes to complete their writing operations independently and quickly, then the segments are assembled in the background, thus improving computing efficiency while maintaining restoration simplicity.
3Reliability
If checkpoint files are written synchronously to the parallel file system, then data consistency is ensured, but the time required for checkpointing increases
Solution Approach 1:
The patent segments the synchronous write operation into asynchronous local writes to burst buffers. Each compute node writes its segment locally without waiting for others, then the segments are assembled in the background. This maintains data consistency through proper synchronization mechanisms while dramatically reducing the perceived checkpointing time, thus resolving the contradiction between reliability and time loss.
Solution Approach 2:
The patent enables continuous useful action by allowing compute nodes to write checkpoint segments asynchronously to burst buffers without waiting for the entire checkpointing process to complete. The assembly of segments into the shared file occurs in the background, ensuring that compute nodes can continuously perform useful computation while data consistency is maintained through proper synchronization, thus reducing time loss without sacrificing reliability.
Data Source
AI summary
A data management system and method for accelerating shared file checkpointing. Written application data is aggregated in an application data file created in a local burst buffer memory at a compute node, and an associated data mapping built index to maintain information related to the offsets into a shared file at which segments of the application data is to be stored in a parallel file system, and where in the buffer those segments are located. The node asynchronously transfers a data file containing the application data and the associated data mapping index to a file server for shared file storage. The data management system and method further accelerates shared file checkpointing in which a shared file, together with a map file that specifies how the shared file is to be distributed, is asynchronously transferred to local burst buffer memories at the nodes to accelerate reading of the shared file.