Open nicktrav opened 2 years ago
cc: @itsbilal
I think we could use some more microbenchmarks in engineccl as well.
As a part of this work, I think we should investigate how performance degrades as the data grows to no longer fit in the block cache. Misses suffer the overhead of decryption.
We'd like to quantify the implications of enabling encrpytion-at-rest (EAR) on a cluster, and identifying potential areas for improvement, if material differences are found.
We have various benchmarks that track the performance of EAR, implemented as variants of existing roachperf benchmarks (e.g. KV0, KV95, etc.), that could be used, though it is finicky to easily identify performance differences between the unencrypted vs. encrypted cases.
Some of the existing benchmarks for EAR (e.g.
kv0/enc=true/nodes=1/cpu=32
on AWS or on GCP), do not appear to run, relative to theirenc=false
counterparts. We should fix / prune these while we're in here quantifying the impacts of EAR.Jira issue: CRDB-15438