2019 Theses Doctoral
Design Space Exploration of Accelerators for Warehouse Scale Computing
With Moore’s law grinding to a halt, accelerators are one of the ways that new silicon can improve performance, and they are already a key component in modern datacenters. Accelerators are integrated circuits that implement parts of an application with the objective of higher energy efficiency compared to execution on a standard general purpose CPU. Many accelerators can target any particular workload, generally with a wide range of performance, and costs such as area or power. Exploring these design choices, called Design Space Exploration (DSE), is a crucial step in trying to find the most efficient accelerator design, the one that produces the largest reduction of the total cost of ownership.
This work aims to improve this design space exploration phase for accelerators and to avoid pitfalls in the process. This dissertation supports the thesis that early design choices – including the level of specialization – are critical for accelerator development and therefore require benchmarks reflective of production workloads. We present three studies that support this thesis. First, we show how to benchmark datacenter applications by creating a benchmark for large video sharing infrastructures. Then, we present two studies focused on accelerators for analytical query processing. The first is an analysis on the impact of Network on Chip specialization while the second analyses the impact of the level of specialization.
The first part of this dissertation introduces vbench: a video transcoding benchmark tailored to the growing video-as-a-service market. Video transcoding is not accurately represented in current computer architecture benchmarks such as SPEC or PARSEC. Despite posing a big computational burden for cloud video providers, such as YouTube and Facebook, it is not included in cloud benchmarks such as CloudSuite. Using vbench, we found that the microarchitectural profile of video transcoding is highly dependent on the input video, that SIMD extensions provide limited benefits, and that commercial hardware transcoders impose tradeoffs that are not ideal for cloud video providers. Our benchmark should spur architectural innovations for this critical workload. This work shows how to benchmark a real world warehouse scale application and the possible pitfalls in case of a mischaracterization.
When considering accelerators for the different, but no less important, application of analytical query processing, design space exploration plays a critical role. We analyzed the Q100, a class of accelerators for this application domain, using TPC-H as the reference benchmark. We found that the hardware computational blocks have to be tailored to the requirements of the application, but also the Network on Chip (NoC) can be specialized. We developed an algorithm capable of producing more effective Q100 designs by tailoring the NoC to the communication requirements of the system. Our algorithm is capable of producing designs that are Pareto optimal compared to standard NoC topologies. This shows how NoC specialization is highly effective for accelerators and it should be an integral part of design space exploration for large accelerators’ designs.
The third part of this dissertation analyzes the impact of the level of specialization, e.g. using an ASIC or Coarse Grain Reconfigurable Architecture (CGRA) implementation, on an accelerator performance. We developed a CGRA architecture capable of executing SQL query plans. We compare this architecture against Q100, an ASIC that targets the same class of workloads. Despite being less specialized, this programmable architecture shows comparable performance to the Q100 given an area and power budget. Resource usage explains this counterintuitive result, since a well programmed, homogeneous array of resources is able to more effectively harness silicon for the workload at hand. This suggests that a balanced accelerator research portfolio must include alternative programmable architectures – and their software stacks.
- Lottarini_columbia_0054D_15082.pdf application/pdf 2.96 MB Download File
More About This Work
- Academic Units
- Computer Science
- Thesis Advisors
- Kim, Martha Allen
- Ph.D., Columbia University
- Published Here
- February 7, 2019