Lemmy Benchmarking

Rust code that instruments lemmy server to check for performance.

Uses PostgreSQL pg_stat_statements to get query counts and perf to build flame graphs.

Uses deno+react to render nice markdown/html files of the results. Results are uploaded to https://github.com/phiresky/lemmy-perf-data/tree/main/comparisons (currently manually).

Implemented Benchmarks

Import-Reddit-Dump

The benchmark Import-Reddit-Dump-v1 reads a real dump of reddit content from December 2022, limited to 100k events, and uploads it to a lemmy server as if reddit was another federated instance. The votes use the real counts but simulated users and times (uniformly randomly between 0 and 2h after the post/comment).

Todo:

• add warmup phase to improve comment/post vote/c, vote/p ratios.
• log memory usage while benchmarking (getrusage or wait3?)

Todo: user-facing benchmark

Simulate user load (posting, commenting, upvoting)

Notes

how much activity is there actually?

with start as (select id, published from activity where id > (select max(id) from activity) - 100000 order by id limit 1) select now(), (select published as start from start), count(activity), sum(length(activity.data::text)) as bytes from activity where activity.id >= (select id from start);

lemmy.world:

              now              |           start            | count |  bytes
-------------------------------+----------------------------+-------+----------
 2023-07-06 12:50:31.136366+00 | 2023-07-06 11:16:00.563352 | 93709 | 80491448

that's around 50MB/hour (16.6 activities per second)

lemm.ee:

now,start,count,bytes
2023-07-06 11:38:24.840290 +00:00,2023-07-06 09:43:28.282779,96912,54744286

that's around 30MB/hour (14 activities per second)