|
Wensong Zhang wrote:
>
> Actually, there was a joint demostration of
> LVS/ldirectd/heartbeat/lvs-gui package in the LinuxWorld Expo. Horms
> told me:
> "As I wasn't at the show I am still picking up reports of
> what happened but I understand that there was some
> heavy connections/second testing conducted on the 16 web server
> farm that was put together and it made Turbo Linux - in the booth
> opposite - look a little lean."
>
> ;-)
>
> Wensong
Hello guys,
Since I was probably the one that cited some internal benchmark
statistics, I believe it's probably up to me to clarify the
configuration. I'll preface this with the statement that our goal was
not to hammer TurboCluster Server as hard as we possibly could. We were
simply looking what it would take to achieve a reasonable value.
Internally, we've achieved 20,000 connections/s using our routing
software. The configuration was 2 Celeron-based servers and 2
Celeron-based clients. Additionally, we were using a stock untuned Linux
2.2.13 kernel. In our benchmark, the most significant bottleneck we
identified was the Linux kernel itself.
The individuals I spoke with at VA Linux claimed 50,000 connections/s.
If the above configuration setup refers to this value, it looks like 16
real servers were required to generate this traffic. As VA makes
top-notch excellent hardware, I suspect they were using higher end
servers than our small Celeron systems. The VA engineers also mentioned
that they were using a modified 2.2 or 2.3 kernel, with improved TCP/IP
performance characteristics.
All of these numbers are relatively meaningless, however, without a
standardized benchmark framework. As any benchmarker knows, any number
can be achieved if you modify the test harness and setup sufficiently.
:)
In the spirit of good healthy fun, perhaps it would be interesting to
have a benchmark "cook-off" one of these days? We could either do this
as an engineering project, just for the fun and curiosity of the
participants, or as something more formal. Regardless of who comes out
on top, it would probably be great information for each of our
respective engineering teams to take back and use to further refine each
of our products. To keep it friendly, I propose that the losing team buy
the winning team a round of beer. :)
This would probably be most convenient for those of us in the Bay Area,
although I'd be happy to tell our Beijing office to take Wensong Zhang
out for a mao tai. :)
Best Regards,
Aaron
--
Aaron McKee
Sr. Clustering Products Manager
TurboLinux Inc.
aaron.vcf
Description: Card for Aaron McKee
|
