karczex
commented
3 years ago Basic scope:
Key/value sizes: 8/8, 128/128
Benchmarks: fillseq, fillrandom readrandom, readrandom (after fillseq) readrandom, readrandom (after fillrandom) readseq, readseq (after fillseq) readseq, readseq (after fillrandom) readwhilewriting readrandomwriterandom
Number of elements: 100 000 000
Number of threads (for concurrent engines): 1, 4, 8, 12, 16, 18, 24
Number of dimms: 6 interleaved
Medium:
- pmem (fsdax)
- DRAM (/dev/shm with PMEM_IS_PMEM_FORCE=1)
Engines: cmap, csmap, radix, stree and robinhood (when merged)
Ideally we would test all the combinations. After analysis we can reduce the numbers to the most representative ones.
Metrics we should collect: bandwidth, ops/sec, latency
Key/value sizes: 8/8, 24/24, 8/128, 128/128, 256/256,
Benchmarks: fillseq, fillrandom readseq, readseq (after fillseq)
readseq, readseq (after fillrandom) readrandom, readrandom (after fillseq) readrandom, readrandom (after fillrandom) overwrite, overwrite (after fillseq) overwrite, overwrite (after fillrandom) readmissing, readmissing (after filrandom or fillseq) deleteseq (after fillseq) deleteseq (after fillrandom) deleterandom (after fillseq) deleterandom (after fillsrandom) readwhilewriting
readwhilewriting (after fillrandom) readwhilewriting (after fillseq) readrandomwriterandom readrandomwriterandom (after fillrandom) readrandomwriterandom (after fillseq)
Number of elements: 1 000 000, 10 000 000, 100 000 000
Number of threads (for concurrent engines): 1, 2, 4, 6, 8, 12, 18, 24
Flags: disjoint=1 disjoint=0
Number of dimms: 1, 6 interleaved
Numa node:
OPTIONAL (second step):
Pool utilization (TODO: think how to approximate the pool size): 10%, 50%, 90%
fsdax/devdax
compilers: clang++ g++