On 10/15/2009 02:38 AM, jean-frederic clere wrote:
=
Which kind of hardware?
8-core Intel @1.86GHz (a few years old stuff). Most of the tests /
benchmarks don't become CPU bound though. Some become network bound,
some (we suspect) hits limitations in our kernel (all our tests are on a
very old kernel, RHEL4 with 2.6.9), and some might even become limited
by SMP issues (e.g. poor utilization of cache lines).
This is definitely something we'll look forward working on with the
community. I firmly believe we have several ways of improving
performance and scalability, e.g.
1. Take advantage of new features and improvements in modern kernels.
2. Take advantage of new libraries, tools and methodologies, to
improve scalability on 8 and more CPUs
3. Investigate where our true bottlenecks are on modern HW and OSes.
For example, we've spent little time optimizing L1/L2 cache
utilization, isolating and eliminating false sharing etc.
I'm pretty convinced our design / architecture can drive traffic at much
higher rates that what we see, we just need to get better at using
modern HW and software :).
Cheers,
-- Leif
