Re: hand-wave-y cache performance data

On 1/12/2012 10:06 AM, Ehsan Akhgari wrote:
> On Tue, Jan 10, 2012 at 10:13 PM, Brian Smith <[email protected] 
> <mailto:[email protected]>> wrote:
>
>     Boris Zbarsky wrote:
>     > On 1/10/12 8:26 PM, Jason Duell wrote:
>     > > The take-away would seem to be that we're not doing anything
>     really
>     > > horrible (yay), and are competitive with IE/Opera, but we could
>     > > improve somewhat (Chrome tends to be faster).  Nothing earth-
>     > > shattering or market-defining, IMO.
>     >
>     > The main take-away for me was that for uncached loads of big complex
>     > pages we do pretty comparably to Chrome, but for cached ones we're
>     > way slower....
>
>     This is how I interpreted it too.
>
>     I think also that Taras's concern is more about startup time
>     issues with the cache, with more of an emphasis on worse-case
>     performance as opposed to the best case (good hardware)
>     configurations like this test seems to be emphasizing.
>
>
> Do you guys think that it would be useful to get somebody to benchmark 
> how much overhead our cache introduces?  Doing this much worse with 
> the cache enabled is bad IMO.  A simple first step would be for 
> somebody to submit Talos runs on a try server job with the disk (and 
> maybe mem) cache disabled and compare some numbers...
QA is doing some manual testing comparing us to other browsers, etc. See 
https://wiki.mozilla.org/Performance/Snappy/Testing:BFCache_Sprint#Reference_Test_Case 
. We will make this more complex/concrete in future iterations by adding 
xperf, wireshark into the mix.
My concern isn't worst case perf specifically. I'm concerned that a 
cache could actually be lagging us.  We should change our implementation 
to make that impossible(ie 
https://bugzilla.mozilla.org/show_bug.cgi?id=715714). I don't care that 
other browsers have an even crappier cache implementation than we do :)

Taras