Pebble Benchmarks

Last updated

Benchmarks are run nightly using pebble bench ycsb on AWS m6id.4xlarge machines equipped with local SSD storage. The AWS instances show remarkably high instance to instance performance variability. In order to smooth out that variability the benchmarks are run multiple times each (using different instances) and outliers are excluded.

Detail: Bytes Read | Bytes Written | Read Amp | Write Amp

Options: Local scale

L0-sublevels and flush-splits enabled

Increased LogWriter free blocks 4->16

Began tracking ycsb/E read-amp

Level metadata switched to use a B-Tree

Enabled read-triggered compactions

Readahead and preallocation bug fixed

Removed excess read samples for read-triggered compactions

Switched to Ubuntu 20.04.2 LTS AMI

Read compaction fixes

Bumped benchmark runtime to 90 minutes

Data quality issue introduced (YCSB A only)

Data quality issue fixed (YCSB A only)

Began zeroing reused iterator structs (#1822)

Grandparent boundary compaction splitting

Infrastructure change (#2578)

ycsb/F sampling bug

Switched to m6id.4xlarge (from 5d.4xlarge)

YCSB A (50% reads, 50% updates, zipf key distribution)

YCSB B (95 reads, 5% updates, zipf key distribution)

YCSB C (100% reads, zipf key distribution)

YCSB D (95% reads, 5% updates, uniform key distribution)

YCSB E (95% scans, 5% updates, zipf key distribution)

Insert-only (100% inserts, zipf key distribution)

Write throughput (100% inserts, zipf key distribution)

This benchmark attempts to find the optimal write throughput by driving more and more load against the DB until a target heuristic fails (currently a mixture of number of L0 sublevels, L0 files, and whether the DB has experienced a write stall). These benchmarks are run nightly using pebble bench write on GCP n2-standard-32 machines equipped with 16 local SSDs in a RAID 0 array.