nilhoel1 / bhive

BHive: A Benchmark Suite and Measurement Framework for Validating x86-64 Basic Block Performance Models

More information about our tool/benchmark can be found in this paper.

• BHive: A Benchmark Suite and Measurement Framework for Validating x86-64 Basic Block Performance Models
  Yishen Chen, Ajay Brahmakshatriya, Charith Mendis, Alex Renda, Eric Atkinson, Ondrej Sykora, Saman Amarasinghe, and Michael Carbin
  2019 IEEE International Symposium on Workload Characterization
  

@inproceedings{bhive,
  title={BHive: A Benchmark Suite and Measurement Framework for Validating x86-64 Basic Block Performance Models},
  author={Chen, Yishen and Brahmakshatriya, Ajay and Mendis,  Charith and Renda, Alex and Atkinson, Eric and Sykora, Ondrej and Amarasinghe, Saman and Carbin, Michael},
  booktitle={2019 IEEE international symposium on workload characterization (IISWC)},
  year={2019},
  organization={IEEE}
}

What's here?

Benchmark

• benchmark/categories.csv lists all the basic blocks and their categories.
• benchmark/throughput/ contains the measured throughput of basic blocks. The throughput of some basic blocks are missing because we can't get a clean measurement.
• benchmark/sources/ lists the source applications of all the basic blocks. For each basic block, we additionally show the (static) frequency with which it shows up in an application.
• benchmark/disasm is a tool that disassembles hex representation of a basic block. Use it like this ./disasm 85c044897c2460. It uses llvm-mc and assumes that you have llvm installed.

Basic Block Profiler

timing-harness/ is a profiler that allows you to profile arbitrary, memory accessing basic blocks such as those in sources/.

Throughput Calculation

• The unit of the throughput numbers in the directory benchmark/throughput* is cycles per hundred-iterations.
• The profiler in timing-harness only produces an unrolled basic block's latency---including measurement overhead. To get the final throughput, numbers in benchmark/throughput* is calculated as (L_a-L_b)/(a-b), where a and b are two integer unroll factors (a > b) and L_a, L_b are the the latency of the basic block unrolled a and b times respectively.
• Measurements from the profiler can have noise. The published paper uses the minimum of the measured latencies as the final latency of an unrolled basic block, and the numbers in the master branch are produced using the same methodology.
• The profiler sometimes produces small, spurious latency. Using the median is more stable in this case, and the throughputs are retabulated in a separate branch.

Page Aliasing

Our profiling methodology relies on mapping potentially large number of virtual pages to a small set of physical pages. This is problematic when two memory accesses access different virtual addresses alised to the same physical ones, creating spurious memory dependence and slow down the basic block unnecessarily. We haven't been able to use hardware counters to detect page aliasing reliably, so we tracked the trace of loads and stores occured during profiling and conservatively mark basic blocks that could have been affected. These basic blocks are listed in benchmark/may-alias.csv.