Multi-Resource Packing for Cluster Schedulers

Authors: Robert Grandl, Ganesh Anathanarayanan, Srikanth Kandula, Sriram Rao, Aditya Akella
Venue: SIGCOMM 2014

Cluster level scheduling is a complex topic in which performance, fairness, and hard constraints must all be considered. Fundamentally, a perfectly fair solution sacrifices performance. This work presents a resource-aware cluster scheduling scheme which maximizes performance and includes additional parameters to balance fairness requirements. For simplicity, I will divide the discussion into two sections: the central idea and additional heuristics.

Tetris performs scheduling by analyzing jobs resource requirements in terms of CPUs, memory, disk I/O, and network usage. Each job, task (a subset of a job), and machine is assigned a resource vector. To determine the optimal positioning of a task, a heuristic is used which takes the dot product of the job's resource requirements vs a candidates available resources. The machine with the maximum dot product is selected to place the job. To incorporate fairness into the approach, the number of candidate jobs can be adjusted. Maximizing fairness will only place jobs in the global queue order, while maximizing performance will consider all jobs in the queue.

In addition to the basic setup, Tetris also includes additional heuristic properties to ensure no low-hanging fruit is left on the table. Firstly, shortest remaining time first (SRFT) algorithm is incorporated into the scheduling calculations, such that dot products are weighted to favor jobs nearly completions. The next optimization comes from the nature of map-reduce and DAG dependencies. Again, when a job is nearing the end of a stage, it's tasks a favored so that it can begin the next stage sooner. The last significant optimization is regarding the "remote penalty". When a job is scheduled away from data locality, it's network requirements will increase. Thus, jobs which cause this are assigned a penalty (note this part is a bit vague in the details from my reading).

Overall, exceptional gains are achieved of a 40% improvement in job completion time and 41% improvement in job makespan (the total time for all jobs to complete). This occurs with a minimal impact in fairness, where about 6% jobs are delayed more than 10%. It would have been nice to see an S-curve of job improvements to illustrate the details more clearly, but nevertheless fairness seems to suffer marginally.


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