Hacker News Reader

Comment by jeffbee | Hacker News Reader

Anyone able to replicate this? With XFS on an NVMe device I only get 500 inserts per second. I wonder if the 15000 figure is attributable to the scaleway virtual server SCSI device ignoring syncs.

View on HN

They're using WAL and synchronous = NORMAL, so there shouldn't be that many syncs to begin with.

By default, I think it'd sync only on checkpoints, and a checkpoint would happen for every 1000 pages of WAL. 1000 is the default value of the wal_autocheckpoint pragma.

I am very surprised that you cannot reproduce more than 500 inserts/second on an NVMe drive with the same settings. I didn't try running this benchmark, as I'm not familiar with rust, but I've achieved ~15k/sec write transactions with python bindings under similar settings.

Using the code from the repo, I got this on a standard Dell XPS laptop, running Ubuntu 20.04.1 with ext4 on an NVMe:

    $ cargo run --release -- -c 100 -i 100000
    ...
    Inserting 100000 records. concurrency: 100
    Time elapsed to insert 100000 records: 7.939734809s (12594.88 inserts/s)

Well, there's something about either XFS, Linux 6.2, or the sqlite libraries on Ubuntu 23.04 because on my machine it's hitting 300 flushes per second while doing about 1300 writes / 13MB per second.

> pragma temp_store = memory;

Did you do this?

I'm probably missing something, but why would the temp_store pragma be relevant? They don't appear to be using temp tables in this benchmark.

I am just taking their repo and command directly from their blog post.

Reply on news.ycombinator.com