On 10/08/2009 09:25 AM, jean-frederic clere wrote:
Hi all,
Is there somewhere of benchmark of the exist code?
We haven't produced any 'official' benchmark for the OpenSource release
yet, partially because we don't know what the numbers will be for the
current "open source branch" (we've done a lot of code cleanup changes
which might affect performance a bit). But I can give some ballpark numbers:
Serving small content out of cache: in the 30-40,000 QPS if there's
some keep-alive efficiency
Served out of cache, but no keep-alive: 18-20,000 QPS (and CPS,
obviously)
No cache, proxying everything: 15-18,000 QPS with keep-alive to UA
(not to origin)
No cache, proxying and no keep-alive: 10-15,000 QPS
For "typical" Web traffic, where objects are in the 32kB range (or
more), we should become network bound on all modern HW. Y! internally
deploys all Traffic-Server instances with at least a dual-NIC (bonded).
In lab, we've driven traffic all the way up to 4x NICs (3.6GBps to be
exact).
Once we get this opensourced (soon now ...), and we can spend some time
tuning this for various modern Linux kernels, we can reach the above,
and more. Getting more people to help isolating bottlenecks, contention,
false sharing etc. will be great.
-- leif
