Re: [PATCH] Put sha1dc on a diet

From: Jeff King <peff@peff.net>
Date: Mar 1, 2017, 19:53 UTC
Message-ID: <20170301195302.3pybakmjqztosohj@sigill.intra.peff.net>
In-Reply-To: <CA+55aFx1wAS-nHS2awuW2waX=cvig4UoZqmN5H3v93yDE7ukyQ@mail.gmail.com>

On Wed, Mar 01, 2017 at 10:49:55AM -0800, Linus Torvalds wrote:

> That said, I think that it would be lovely to just default to
> USE_SHA1DC and just put the whole attack behind us. Yes, it's slower.
> No, it doesn't really seem to matter that much in practice.

My biggest concern is the index-pack operation. Try this:

  time git clone --no-local --bare linux tmp.git

with and without USE_SHA1DC. I get:

  [w/ openssl]
  real	1m52.307s
  user	2m47.928s
  sys	0m14.992s

  [w/ sha1dc]
  real	3m4.043s
  user	6m16.412s
  sys	0m13.772s

That's real latency the user will see. It's hard to break it down, though. The actual "receiving" phase is generally going to be network bound. The delta-resolution that happens afterwards is totally local and CPU-bound (but does run in parallel).

And of course this repository tends to the larger side (though certainly there are bigger ones), and you only feel the pain on clone or when doing an initial push, not day-to-day.

So maybe we just suck it up and accept that it's a bit slower.

-Peff