On Fri, Jul 28, 2023 at 10:40:48AM +0200, Omar Polo wrote:
> we're pretty slow at deltifying things.  it's OK, git itself is bad
> with big files, but we can be too slow with moderately-sized files.  I
> noticed this when a couple of days ago I hit the gotd timeout trying
> to push some changes on a private repository because got was taking
> too long to deltify ~100mb of data splitted mostly in 3-4 files.
> 
> (The good news is that gotd timeout works!)
> 
> Most of this slowness comes from resizing the table.  Performance
> degrades when the table is mostly filled, and by incrementing its size
> by 64 slot every time we end up reallocating it quite a lot for big
> files.
> 
> There's also the memory usage to keep low though, as this code is also
> used in gotd, we don't want the table to become 'too big' by
> e.g. doubling its size at every resize.  Note however that there's
> nothing preventing the table to grow arbitrarly large, it just takes
> forever and that we don't have cancellation in place in this code path.
> 
> I've set up a repository to do some tests: it's a standalone binary
> that wraps the deltify code: <https://codeberg.org/op/deltify/>.
> 
> The main branch is the exact same code we have in got, the branch
> `256' bumps the resize step from 64 to 256.
> 
> Just as-is the improvements are noticeable; on my hw deltifying 20M of
> /dev/random takes around 8 seconds, when before it was at 75 seconds.
> 
> Nothing stops us from using an even bigger step size (with 512
> deltifying 20M takes ~2.5 seconds), but since we're already pretty
> conservative in the amount of memory used and I'm one of the first to
> have hit this I decided to stay on the low side.
> 
> The memory usage doesn't grow considerably: the entries of the table
> are 24 bytes long on amd64 (two off_t and one uint32_t, plus padding),
> so in the worst case we end up having allocate 24 * 256 = ~6k of
> memory.
> 
> Future work on this front includes adding cancellation support.
> 
> thoughts?

Please put this in with my ok. We'll see if anyone runs into issues.

I don't think we can make very educated guesses about the optimal value.
We need to keep experimenting until we feel that we've found a threshold
that works, and/or maybe change the scaling strategy eventually.