spa

 ____  ____   _    
/ ___||  _ \ / \   
\___ \| |_) / _ \  
 ___) |  __/ ___ \ 
|____/|_| /_/   \_\ -- Stall-based Performance Analysis

Description

The key insight of SPA is that drilling down the differential CPU stalls between slow and fast memory can yield accurate performance slowdown analysis whereas standalone setup cannot. SPA aims to pinpoint the specific stall sources that contribute to slowdowns, bridging the gap between architectural level and workload-level performance.

The memory performance slowdown is composed of several sections, which are defined in the paper: (1) DRAM slowdown, (2) Cache slowdown (L1/L2/L3), (3) Store slowdown, (4) Core slowdown, and (5) Other. The slowdown is mainly dominated by (1) DRAM slowdown, (2) Cache slowdown (L1/L2/L3), (3) Store slowdown, while (4) and (5) have minimal contribution on the overall slowdown.

Usage

It measures the performance counters used in the paper. The output files are in rst for each workload suite. The data processing files are provided in proc.

The use cases demonstrated in the paper are mainly related to CXL/remote memory performance analysis. Howerer, I believe the analysis methodology should not be limited to CXL/remote memory. Welcome to find out more use cases for this method.

Steps

• Setup (Install packages and perf)

  ./setup.sh
  

• Go to a workload suite directory, (for example, cpu2017), run the 1st workload in the suite:

  sudo ./run.sh w.txt 1
  

  Run all workloads:

  sudo ./run.sh w.txt
  

  Set PERF_INTERVAL_MODE=0 in config.sh to save one aggregate count per run in rst/aggregate. To sample the counters over time, set PERF_INTERVAL_MODE=1 and configure the sampling period with PERF_INTERVAL_MS (milliseconds). The time-series results are saved in rst/interval.

• Collect the performance counters from the output files. proc provides a basic processing code for visualizing the slowdown breakdown.

  • Copy the rst to the directory where update_data.py and process.py is.
  • Generate data files in csv

  python3 update_data.py
  

  Process only one mode when needed:

  python3 update_data.py aggregate
  python3 update_data.py interval
  

  • Process the data in csv and generate plots in plots

  python3 process.py
  

  Process only one mode when needed:

  python3 process.py aggregate
  python3 process.py interval
  

  The aggregate plot is generated in plots/aggregate. Interval plots are generated for each workload in plots/interval; use --bins to set the number of instruction intervals (default: 200):

  python3 process.py interval --bins 100
  

Notes

• The run.sh files run each workload once on local (NUMA node 0) and once on remote (NUMA node 1). The default remote NUMA node is set as 1. For the remote memory other than NUMA node 1 in multi-nodes servers, set --membind 1 to other values.