Skip to content

benchmarking.md

Latest commit

 

History

History
82 lines (56 loc) · 2.11 KB

benchmarking.md

File metadata and controls

82 lines (56 loc) · 2.11 KB
Raw

Benchmarking

Tools and recipes for measuring PrisKV performance. For a quick smoke test, see the README. For server tuning, see Server Configuration.

Table of Contents

priskv-benchmark

The client ships with a benchmark tool that drives load against a running server.

./client/priskv-benchmark -a <server-address> -p 18512 \
    -o set -e 1 -k 16 -v 4096 -t 20

Common flags:

Flag Description
-a ADDR Server address.
-p PORT Server port (default 18512).
-o OP Operation: set, get, etc.
-k BYTES Key length.
-v BYTES Value length.
-t SECS Test duration.
-e N Number of inflight requests per client.

Run ./client/priskv-benchmark -h for the full list.

CUDA / GPU Direct RDMA

Requires building with PRISKV_USE_CUDA=1:

make PRISKV_USE_CUDA=1

The repo ships an end-to-end CUDA/GDR example. See client/example.c for the full source covering:

  • client initialization and connection
  • CUDA integration for GPU Direct RDMA
  • device and host buffer management
  • asynchronous operations driven by epoll

Run it:

./client/priskv-example

Ascend NPU

Requires building with PRISKV_USE_ACL=1:

make PRISKV_USE_ACL=1

Then run the standard benchmark binary; NPU-resident buffers are used when available:

./client/priskv-benchmark -a <server-address> -p 18512 \
    -o set -e 1 -k 16 -v 4096 -t 20

Comparison vs Valkey

Build the bundled valkey-benchmark for a side-by-side comparison:

git submodule init && git submodule update
cd client && make valkey-benchmark

./valkey-benchmark -a 127.0.0.1 -p 6379 -o set -e 1 -k 16 -v 4096 -t 20

Match the flags (-k, -v, -e, -t) when comparing to priskv-benchmark so the two runs exercise comparable workloads.