Low-overhead Remote Persistence for Scalable Low-latency File Systems
TimeTuesday, June 23rd4:40pm - 4:45pm
DescriptionUtilizing emerging persistent memory (PMEM) devices is required to improve the performance of parallel file systems. Parallel file systems use remote I/O operations to distribute data blocks to different storage nodes. However, the conventional remote I/O stack has extra software layers and incurs additional copy between multiple buffer memories. As a result, the latency of remote write operations is dominated by software overhead and the system cannot leverage the nature of low-latency PMEM devices. This poster describes the optimized remote persistence model, which can reduce the latency of remote write operations with persistency by reducing the number of memory copy operations and the number of intermediate software layers. We implemented the model with InfiniBand Verbs API and Intel DC Persistent Memory (DCPM) devices and evaluated it with remote write RPC operations. In order to ensure the durability of written data blocks to PMEM devices, CLWB (Cache Line Write Back) and SFENCE operations are used every time a remote write operation was issued. For the comparison, we used DCPM as a block device and formatted it as an XFS file system partition. The evaluation results show that the optimized model can reduce the latency of the remote write operation by 78% and achieved 11 times faster throughput than the baseline (XFS) case. The optimized model outperforms the conventional remote I/O mechanism, although the both cases use the same hardware. Future work is implementing a complete set of I/O operations by using the optimized persistent mechanism.