Thread Cluster Memory Scheduling: Exploiting Differences in Memory Access Behavior

Authors: Yoongu Kim, Michael Papamichael, Onur Mutlu, and Mor Hachol-Balter
Venue:   MICRO 2010

This paper presents Thread Cluster Memory Scheduling (TCM), a memory scheduling algorithm that targets optimizing both system throughput and fairness. To achieve this, three key ideas are employed.

Firstly, threads are clustered as either bandwidth intensive or non-intensive. The idea here is that low-bandwidth threads are more sensitive to latency. While an example is presented, an easier way to reason about this is that a low-bandwidth thread needs only a small fraction of memory service time to see a significant performance increase. In other words, it has a high ROI with minimal impact to other threads. As such, non-BW-intensive threads are always given the highest priority.

The second observation to be exploited is the disparity in behavior among high-bandwidth threads. Specifically, a metric niceness measures a threads susceptibility to interference and impact on other threads. Threads with high row-buffer locality will starve of other threads, thus threads with high RBL are "hostile" and have a lower nice value. Similarly, threads with high bank level parallelism is more susceptible to hostile threads, and thus given a higher nice value.

The final piece of the puzzle is in the scheduling algorithm, Insertion Shuffle, which employs niceness values. Insertion shuffle minimizes the amount of time hostile threads are given priority and maximizes the time nice threads given priority. There is an interesting auxiliary effect which the authors refer to as "memory service leaks", which allows nice threads of high priority actually allow lower priority threads to be serviced. Finally, in the event that all threads have relatively similar behavior, a randomized shuffle is used.

In all, TCM provides more throughput and higher fairness compared to existing state-of-the-art such as FR-FCFS, STFM, PAR_BS, and ATLAS. In addition, TCM's thresholds allow for a predictable trade-off between throughput and fairness. The separate clusters also allow for OS thread prioritization (weighting) without losing complete fairness by only applying prioritization in relation to the thread's respective cluster.

Full Text

Comments

Post a Comment

Popular posts from this blog

Fundamental Latency Trade-offs in Architecting DRAM Caches (Alloy Cache)

Authors: Moinuddin K. Qureshi and Gabriel H. Loh Venue:    MICRO 2012 The authors present a new architecture for DRAM cache which targets optimizing latency rather than hit rate. The authors show that for every optimization made to improve hit rate, one must also consider the added cache latency, thus resulting in a new "break even hit rate". The authors compare to three key baselines: an infeasibly large SRAM-tag array, the prior "LH-Cache", and a ideal latency-optimized cache. They find that "unoptimizations" improve latency, while minimally impacting hit-rate. Their final result is a cache which boosts performance 35% improvement over the baseline, and nearly 20% better than the previous LH-cache. The specific contributions are: Moving to a direct mapped cache provides significant improvement (by reducing latency) Combining tag-store and data-store into a single entity improves performance 21% Implementing a small per-core predictor enhances per...
Read more

ZCOMP: Reducing DNN Cross-Layer Memory Footprint Using Vector Extensions

Authors: Berkin Akin, Zeshan A. Christi, and Alaa R. Alameldeen Venue: MICRO, 2019 With accelerators dominating the deep learning space in architecture conferences, this paper stands out as it focuses on reducing DNN inference/training overhead while using a CPU. An obvious question is, with the wide-spread use of GPUs/TPUs for deep learning, why should we focus on optimizing CPUs for deep learning. In a recent paper published at HPCA 2019, Facebook claims that CPUs are preferred for applications where a tight integration is required between DNN and non-DNN tasks. Also, Intel's recent AVX512 has specialized support for DNNs in the form of new instructions called Vector Neural Network Instructions (VNNI). Broadly, optimizing DNNs can be viewed from two different perspectives; computation, and communication. This paper targets reducing communication overhead, more specifically, activation or feature-map communication overhead by compressing them. Compressing activations/weights...
Read more

AutoFDO: Automatic Feedback-Directed Optimization for Warehouse-Scale Applications

Authors: Dehao Chen, David Xinliang Li, Tipp Moseley Venue:    CGO 2016 This paper presents AutoFDO, a system used which profiles warehouse-scale applications, and applies feedback to the compilation for the next release. AutoFDO works by profiling and storing profiles in an aggregate database, annotating the profiles via an intermediate representation, and finally providing feedback. On average, this technique boosts performance by around 10% and works well even with stale releases. For profiles to be useful, they must be in an intermediate representation. This is built from binary-level profile, which uses LBR to map instruction frequencies. By using these frequencies, program counters, and the source, a source profile can be generated. The source profile can be used to build an annotated call-graph with edge frequencies, with some inaccuracy. These annotated call-graphs can then be used as feedback to the compiler. Because of the size of the applications and scale of d...
Read more