trafficserver-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Reindl Harald <>
Subject Re: 3.3.5: performance regression?
Date Sat, 10 Aug 2013 20:50:02 GMT

Am 10.08.2013 22:24, schrieb Leif Hedstrom:
> On Aug 10, 2013, at 1:08 PM, Reindl Harald <> wrote:
>> Am 10.08.2013 19:47, schrieb Reindl Harald:
>>> i am currently working at my F18 SPEC-file to reflect the latest
>>> changes in my packaging and rebuild 3.2.5 ASAP on Fedora
>>> 19 so that "yum update" and "yum downgrade" gives a better picture
>>> but i doubt that the 3 years old XEON at the company will
>>> outperform the one year old IvyBdrige at home with the
>>> same environment, benchmarks besides ATS are compareable
>>> we will see
>> Am 10.08.2013 19:37, schrieb Leif Hedstrom:
>>> Yeah I've tested it with my normal perf regressions, and 3.3.5 was
>>> the same as 3.2.4. In my small object test I get 160,000 qps
>> on what hardware straight from hell do you get 160 thousand qps?
>> however, back to topic...…
> model name	: Intel(R) Core(TM) i7 CPU         920  @ 2.67GHz
> I believe Bryan has clocked it at close to 300,000 QPS on modern dual socket machines

hmmmm - if i find enough spare time i will test ATS on teh host-setup
instead the VM, not that i ever will use bare-metal servers, but
i am interested :-)

>> ____________________________________________________________________________
>> OK, sorry for the noise, that's why there was a "?" in the subject
>> same virtual machine with trafficserver-3.2.5-3.fc19.20130810.rh.x86_64
>> Total transferred:      94732065 bytes
>> HTML transferred:       64253704 bytes
>> Requests per second:    3206.60 [#/sec] (mean)
>> Time per request:       62.371 [ms] (mean)
>> Time per request:       0.312 [ms] (mean, across all concurrent requests)
>> Transfer rate:          2966.48 [Kbytes/sec] received
> I have no idea at this moment, but I'll double check when I get home. So I understand,
you get 90k QPS with 3.2.4 and 3k QPS with 3.3.5 ?

no, 3k QPS with both on the same VM
it's not the ATS version as thought

90k QPS are impossible here, at least not with a network-stack betweens ATS and "ab"

> One thing, did you try clearing the cache after upgrading from 3.3.5? 

yes, i learned to delete all cache files und on the production machine
with raw disk dd if=/dev/zero of=/dev/sdc bs=1M count=50 before start
with a changed version

> As for virtualization, I measure around 5-10x worse performance on virtualization vs
bare metal.
> I've tested this on many setups (including AWS), and I've verified it on ATS, Varnish
and Nginx
> (they all see the same degradation).

we are 100% virtualized for a lot of reasons and it seems the difference here is between
VMware ESXi (bare-metal hypervisor) and VMware Workstation (hosted virtualization) which
in no tests execpt http-proxy made such a big difference over years

most likely the virtual network bare metal versus bridged on a Linux host
however, i am still impressed about the difference and the decision to take
money in the hands for the vSphere cluster seems to be once more confirmed

View raw message