Open asfimport opened 4 years ago
Varun Thacker (@vthacker) (migrated from JIRA)
When we tested C5.4xlarge vs C4.4xlarge the C5s were able to serve \~31% more requests per second. These numbers aren't from running lucene however the C5s might be faster than the C4s ?
Michael McCandless (@mikemccand) (migrated from JIRA)
These numbers aren't from running lucene however the C5s might be faster than the C4s ?
Oh yes the c5
instances are definitely faster than c4
.
For a more fair comparison, I'll try to baseline against a m5.8xlarge
instance instead – both are "general purpose".
Eventually we will have c6g
instances (Graviton2, compute optimized) and then we can compare e.g. c5
and c6
.
Really I'm just curious how Lucene's benchmarks perform on ARM :)
At AWS re:Invent 2019 last week, AWS announced new EC2 instances based on the Graviton2 ARM processor which apparently can be much faster than the original A1 instances, at least according to internal benchmarks.
I've been running Lucene's benchmarks (
wikimediumall
, indexing 33.3 M docs and running a diverse and repeatable set of search tasks) on these instances, comparing ac4.8xlarge
(x86-64) instance against the newm6g.8xlarge
(ARM), and I'll summarize the results here.Net/net ARM seems to be faster at raw indexing than x86-64, even though
m6g.8xlarge
has only 32 cores versus 36 cores, but a bit slower at merging, while searching seems to be faster for some queries and slower for others. I'll try to get the full results posted soon.Migrated from LUCENE-9086 by Michael McCandless (@mikemccand), 1 vote, updated Dec 11 2019