Cache Benchmarks

Summary

Cache Benchmarks These benchmarks measure thoughput, latency, and CPU cycles for Memcache, Redis, Valkey, Dragonfly, and Garnet. Persistance is turned off for all caches, no disk operations. All connections are local, UNIX named pipes. The hardware is an AWS c8g.8xlarge (32 core non-NUMA ARM64). The benchmarking tool is memtier_benchmark. Includes pipelining for 1, 10, 25, and 50. Each benchmark has 31 runs. About 15K total runs. The median of the 31 is used for graphing. Latency is measured in 50th, 90th, 99th, 99.9th, 99.99th percentiles. Latency also includes MAX, the absolute slowest single request. CPU cycles are measured using the perf Linux utility. Linux utility. All graphs are a logarithmic scale. The "Threads" at the bottom of each graph represents the number of I/O Threads that the caching server software is configured to use for that specific benchmark. All caching software benchmarked has some type of multithreaded networking support through the following startup flags. Memcache: -t Redis: --io-threads Valkey: --io-threads Dragonfly: --proactor_threads Garnet: --miniothreads/maxiothreads --minthreads/maxthreads For each benchmark, a fresh instance of the cache server software is started, which is dedicated to 16 cores using taskset -c 0-15 . The memtier_benchmark tool uses the other 16 cores taskset -c 16-31 . Of those 16 cores, there are 256 clients spread evenly between 16 threads. Those clients perform 100K SET and 100K GET operations, each. There's a warmup stage that occurs at the start of each run, just after the cache software is started. It performs a dry run of all SET operations. This warmup is not a part of the measurements. The ./bench-all.sh scripts starts running the benchmarks and produces results that are placed in the results directory. Expect it to take about two weeks from start to finish to complete all runs. CACHE VERSION memcached 1.6.38 Redis v=8.0.2 sha=994bc96b:0 malloc=jemalloc-5.3.0 bits=64 build=29503de25b2919e1 Valkey v=8.1....