Benchmark transaction-submission latency #869

Activity

1. marwen-abid

   added this to Platform Scrumon Jul 17, 2026
2. marwen-abid

   moved this to To Do in Platform Scrumon Jul 17, 2026
3. marwen-abid

   added

   rpc-benchmark

   on Jul 17, 2026
4. Shaptic commented on Jul 17, 2026

   

   Shaptic

   on Jul 17, 2026

   Last edited by Shaptic

   Contributor

   More actions

   Just for context and an initial ballpark, when I analyze the vegeta-provided latency of the mainnet TPS load test (running on an m6id.4xlarge), I see the following:

   Round-trip time of POST /transactions to receiving PENDING, including network latency:

   Workload	Mean
   SAC	67ms
   OZ	59ms
   Soroswap	55ms

   I know this isn't enough to consider this "done," but it is a known starting point with v1 that we can point to for hand-waving if necessary. We should expect the same results, if not better given that network latency isn't part of this metric, with v2 once that's all wired in given that the submission code and execution path should be nearly identical.

   👍React with 👍1Reacted by Marwen Abid
5. marwen-abid

   added

   rpc-full-history

   on Jul 21, 2026
6. marwen-abid

   added this to the platform sprint 74 milestone on Jul 21, 2026
7. marwen-abid

   self-assigned this

   on Jul 21, 2026
8. sydneynotthecity

   added a parent issue on Jul 24, 2026

   • O3.1.2 - Build Infra and Benchmark RPCv2 on Phase 1/2/3 #898
9. marwen-abid

   moved this from To Do to In Progress in Platform Scrumon Jul 24, 2026
10. marwen-abid commented on Jul 28, 2026

    

    marwen-abid

    on Jul 28, 2026

    ContributorAuthor

    More actions

    Results: transaction-submission latency measured inside RPC

    Method. The p99 60 ms submission target decomposes into ~50 ms client↔RPC network RTT plus <10 ms inside the RPC process, so this benchmark measures the in-RPC part and treats the RTT as a constant. Two existing instruments cover it — no new code:

    • finished JSONRPC request log line — full sendTransaction handler duration, one line per request (joined to the starting JSONRPC request line on req to filter by method). This is the headline distribution.
    • soroban_rpc_txsub_submission_duration_seconds — the synchronous RPC→Core POST /tx leg alone.
    • RPC residue (XDR decode, hash, JSONify) = handler mean − core-leg mean. Means only; quantiles do not subtract.

    Environment. stellar/quickstart:testing (stellar-rpc v27.1.1) on an m6id.2xlarge. Load = real fee-bumped Soroban transactions from the tx-load-test tool in stellar-rpc-blaster, three profiles: SAC transfer (≈1 KB), OpenZeppelin token transfer (≈1 KB), Soroswap router swap (≈2.6 KB). Standalone network: 10 RPS × 120 s per profile (~1,150 samples each). Testnet: same box, 2 RPS × 60 s (~115 samples each). Every submission was accepted (PENDING); zero poll errors in every run.

    Standalone (headline):

    Profile	n	p50	p90	p99	max	core-leg mean	RPC residue mean
    SAC transfer	1155	1.14 ms	1.30 ms	4.36 ms	11.4 ms	1.03 ms	0.20 ms
    OZ transfer	1154	1.10 ms	1.25 ms	3.08 ms	15.4 ms	0.99 ms	0.20 ms
    Soroswap swap	1152	1.22 ms	1.39 ms	4.56 ms	13.3 ms	1.07 ms	0.25 ms

    Testnet (confirmation):

    Profile	n	p50	p90	p99	max	core-leg mean	RPC residue mean
    SAC transfer	114	1.16 ms	1.30 ms	4.00 ms	9.2 ms	1.08 ms	0.19 ms
    OZ transfer	115	1.12 ms	1.28 ms	1.43 ms	7.5 ms	0.99 ms	0.20 ms
    Soroswap swap	114	1.19 ms	1.39 ms	1.99 ms	12.6 ms	1.08 ms	0.24 ms

    Findings.

    1. Worst-profile p99 is 4.6 ms — the <10 ms in-RPC slice holds with 2× headroom. Total slice estimate: 50 + 4.6 ≈ 55 ms against the 60 ms target.
    2. The RPC→Core call carries ~80–83 % of the mean and drives the tail (core-leg p99 tracks handler p99 in every profile). RPC's own processing is ~0.2 ms, flat across profiles.
    3. Envelope size barely matters: the ≈2.6 KB Soroswap profile adds ~0.1 ms at p50 over the ≈1 KB profiles.
    4. Standalone ≈ testnet: medians agree within ~2 %, core-leg means within 5 % — the measured path (handler + local core admission) never touches the network. The thinner testnet p99s are sample-size noise at n≈114.
    5. Cross-checks: log-derived means match the soroban_rpc_json_rpc_request_duration_seconds Summary exactly, counts agree across logs, metrics, and the client, and client-observed HTTP RTT (p50 ≈ 2.5 ms) = handler time + the localhost/nginx hop.

    Reproduce. Runner + log-harvest scripts: scripts/tx-submission/ in stellar-rpc-blaster (branch tx-submission-harvest). Raw bundles (client NDJSON, RPC log slices, /metrics scrapes, per-profile summary JSONs): gs://rpc-full-history/results/txsub-standalone-20260728/ and gs://rpc-full-history/results/txsub-testnet-20260728/.

    Measured on v1 (v27.1.1); the v2 submission path is expected to be equivalent — re-verify when it is final.
11. marwen-abid commented on Jul 28, 2026

    

    marwen-abid

    on Jul 28, 2026

    ContributorAuthor

    More actions

    👉 details and graphs https://stellar-experimental.github.io/stellar-rpc-benchmarks/tx-submission.html
12. marwen-abid

    moved this from In Progress to Needs Review in Platform Scrumon Jul 28, 2026
13. marwen-abid

    moved this from Needs Review to Done in Platform Scrumon Jul 30, 2026
14. marwen-abid

    closed this as completedby moving to Done in Platform Scrumon Jul 30, 2026