add use_durable_nonce option #27151
add use_durable_nonce option #27151
Conversation
KirillLykov
force-pushed
the
bench_tps_add_nonce_cli
branch
from
4feeb4a to
dba5f83
Compare
KirillLykov
force-pushed
the
bench_tps_add_nonce_cli
branch
2 times, most recently
from
19de07a to
d4e9d85
Compare
|
@joncinque I've updated PR with |
bench-tps/src/bench.rs
Outdated
| let blockhashes: Vec<Hash> = pubkeys | ||
| .chunks(MAX_MULTIPLE_ACCOUNTS) | ||
| .map(|pubkeys| get_nonce_blockhashes(&client, &pubkeys)) | ||
| .flatten() | ||
| .collect(); |
There was a problem hiding this comment.
nit: Instead of chunking here, it might be faster to pass all of them to get_nonce_blockhashes and then do the chunking in the function, that way you can always fetch MAX_MULTIPLE_ACCOUNTS. What do you think?
There was a problem hiding this comment.
I've updated the code and tested it with private cluster.
I've added UPD section to the PR description with some plots from this run.
It looks like tps is spiky because we request all the blockhashes and later execute all the transactions. Spike width is ~4s while the gap is ~2s => it is possible to increase tps at most by 30% by overlapping these two phases.
I created an issue for this as a follow up since it requires major code reorganization: #27416
There was a problem hiding this comment.
Awesome, and yes it makes sense to do that work separately
KirillLykov
force-pushed
the
bench_tps_add_nonce_cli
branch
from
d4e9d85 to
f03417e
Compare
KirillLykov
force-pushed
the
bench_tps_add_nonce_cli
branch
from
f03417e to
d8a2480
Compare
KirillLykov
force-pushed
the
bench_tps_add_nonce_cli
branch
from
d8a2480 to
bab6632
Compare
thanks! it would be not possible without your comments |
Problem
Final PR in the chain of PRs adding durable nonce transactions to bench-tps.
Requires #27176Requires #27379
Summary of Changes
--use-durable-nonceoption to cliUPD:
TPS is now ~5k
Some plots with running with this option:
The bottom plot is for data shreds: at the beggining there are many data shreds because we create nonce durable accounts. During transactions execution this metric is lower.
If we zoom into the transaction phase, we observe that the tps is spiky:
It happens because it takes time to request blockhashes before executing transactions