FYI: status of svn-fe speed

David Michael Barr· Oct 26, 2011, 15:18 UTC · lore

Hi,

I talked a lot about low-level speed optimisation after the high-level optimisation is finished this week. I merged my svn-fe-pu branch with jch/next and tested with a 1000 rev dump of patches from git-core. When not blocked, svn-fe can process such revs at ~5000 rev/s on my laptop. This is far above what both svnadmin dump and git fast-import presently achieve and might be deemed excessive. However, this translates to low latency which is critical for parallelisation.

I had to use a 10 microsecond sampling interval to get an accurate profile. These numbers have not been normalised but are the complete set of symbols sampled. 2.2% parse_date_basic 1.5% match_string 1.1% svndump_read 1.1% 0x10000af70 [20.2KB] 0.4% strbuf_grow 0.4% strbuf_fread 0.4% strbuf_add 0.4% reset_rev_ctx 0.4% next_quote_pos 0.4% is_date 0.4% fast_export_data 0.4% end_revision 0.4% buffer_read_line 0.4% buffer_copy_bytes

I still think it ought to be a little faster. ;)

-- 
David Barr

Tay Ray Chuan· Oct 26, 2011, 16:18 UTC · re: David Michael Barr · lore

Re: FYI: status of svn-fe speed

On Wed, Oct 26, 2011 at 11:18 PM, David Michael Barr <davidbarr@google.com> wrote:

> I talked a lot about low-level speed optimisation after the high-level
> optimisation is finished this week. I merged my svn-fe-pu branch with
> jch/next and tested with a 1000 rev dump of patches from git-core.
> When not blocked, svn-fe can process such revs at ~5000 rev/s on my

I had to rub my eyes when I read this - what???

Good job, then!

-- 
Cheers,
Ray Chuan