Re: [RFC] On watchman support

From: Duy Nguyen <pclouds@gmail.com>
Date: Nov 28, 2014, 11:13 UTC
Message-ID: <CACsJy8DayFy83JijrkgST5rAbNsst-dgqaP-ebpWXoGKPtp7sA@mail.gmail.com>
In-Reply-To: <1416334360.27401.10.camel@leckie>

On Wed, Nov 19, 2014 at 1:12 AM, David Turner <dturner@twopensource.com> wrote:

Show 11 quoted lines

>> Or will you go
>> with cityhash now.. I ask because you have another sse optimization
>> for hashmap on your watchman branch and that could reduce init time
>> for name-hash. Name-hash is used often on case-insensitive fs (less
>> often on case-sensitive fs).
>
> Cityhash would be better, because it has actual engineering effort put
> into it; what I did on my branch is a hack that happens to work
> decently.  As the comment notes, I did not spend much effort on tuning
> my implementation.  Also, Cityhash doesn't require SSE, so it's more
> portable.

Cityhash looks less appealing to me. For one thing it's C++ so linking to C can't be done. I could add a few "extern "C"" to make it work. But if we plan to support it eventually, cityhash must support C out of the box.

Then cityhash does not support case-insensitive hashing. I had to make a CityHash32i version based on CityHash32. It's probably my bugs there, but performance is worse (~120ms) than original hashmap.c (90ms). Enabling sse4.2 helps a bit, but still worse. Using the case-sensitive version in place for memihash and strihash does make cityhash win over hashmap.c, around 50ms (with or without sse4.2). But that's still not as good as your version (~35ms)..

-- 
Duy