Proceedings of the Second International Workshop on Sustainable Ultrascale Computing Systems (NESUS 2015) Krakow, Poland

Current storage trends dictate placing fast storage devices in all servers and using them as a single distributed storage system. In this converged model where storage and compute resources co-exist in the same server, the role of the network is becoming more important: network overhead is becoming a main limitation to improving storage performance. In our previous work we have designed Tyche, a network protocol for converged storage that bundles multiple 10GigE links transparently and reduces protocol overheads over raw Ethernet without hardware support. However, current technology trends and server consolidation dictates building servers with large amounts of resources (CPU, memory, network, storage). Such servers need to employ Non-Uniform Memory Architectures (NUMA) to scale memory performance. NUMA introduces significant problems with the placement of data and buffers at all software levels. In this paper, we first use Tyche to examine the performance implications of NUMA servers on end-to-end network storage performance. Our results show that NUMA effects have significant negative impact and can reduce throughput by almost 2x on servers with as few as 8 cores (16 hyper-threads). Then, we propose extensions to network protocols that can mitigate this impact. We use information about the location of data, cores, and NICs to properly align data transfers and minimize the impact of NUMA servers. Our design almost entirely eliminates NUMA effects by encapsulating all protocol structures to a “channel” concept and then carefully mapping channels and their resources to NICs and NUMA nodes.