Traffic management

Abstract

NUMA systems are characterized by Non-Uniform Memory Access times, where accessing data in a remote node takes longer than a local access. NUMA hardware has been built since the late 80's, and the operating systems designed for it were optimized for access locality. They co-located memory pages with the threads that accessed them, so as to avoid the cost of remote accesses. Contrary to older systems, modern NUMA hardware has much smaller remote wire delays, and so remote access costs per se are not the main concern for performance, as we discovered in this work. Instead, congestion on memory controllers and interconnects , caused by memory traffic from data-intensive applications, hurts performance a lot more. Because of that, memory placement algorithms must be redesigned to target traffic congestion. This requires an arsenal of techniques that go beyond optimizing locality. In this paper we describe Carrefour , an algorithm that addresses this goal. We implemented Carrefour in Linux and obtained performance improvements of up to 3.6 relative to the default kernel, as well as significant improvements compared to NUMA-aware patchsets available for Linux. Carrefour never hurts performance by more than 4% when memory placement cannot be improved. We present the design of Carrefour , the challenges of implementing it on modern hardware, and draw insights about hardware support that would help optimize system software on future NUMA systems.

Dvé: Improving DRAM Reliability and Performance On-Demand via Coherent Replication

“…We note that the performance of the best variant of the allow protocol for each benchmark is within 12% and 7% of the oracular performance for top10 and all benchmarks (16% better than the baseline NUMA overall). Sensitivity to inter-socket latency: A number of softwarebased techniques -including Carrefour [16], Shoal [35], AutoNUMA -have been proposed to mitigate NUMA effects; the net effect of each of these techniques is to reduce the average inter-socket latency. Therefore, we study the effect of inter-socket interconnect latencies on the performance of Dvé.…”

“…2) Mitigating NUMA overheads: Carrefour [16] proposes OS-driven selective replication of memory read-only or readmostly pages to alleviate NUMA overheads. Shoal [35] proposed program analysis to automatically replicate memory regions across NUMA nodes for mitigating the performance penalty of remote access.…”

EnergAt: Fine-Grained Energy Attribution for Multi-Tenancy

“…We argue that this energy attribution method is insufficient in precisely capturing applications' power dynamics. Specifically, it neither takes into account NUMA effects [8,14,35,55] in the presence of multiple sockets nor does it distinguish threads from processes when tracing energy provenance ( §3.1). This method is also prone to the noisy-neighbor effect [15,16] in a multi-tenant environment, where the energy attribution of an application is interfered by collocated tasks on the same host.…”

A control-inspired iterative algorithm for memory management in NUMA multicores

“…Finally, the work (Dashti et al, 2013) focuses on the memory controller congestion which hampers the performance. In particular, modern NUMA systems reduce the remote access time, while the contention on both memory controllers and interconnect still introduces non-negligible performance penalties.…”

“…In particular, modern NUMA systems reduce the remote access time, while the contention on both memory controllers and interconnect still introduces non-negligible performance penalties. In this respect, the method proposed in (Dashti et al, 2013) does not consider the mutual influence on the memory allocation due to memory-bound tasks, which may force memory swap-out for different applications. Our proposal accounts for this aspect and limits the interconnect congestion promoting local memory placement.…”