|
Polygraph. Check the archives for discussions between me and Joseph
Mack from about 3-4 months ago on the subject of benchmarking LVS. I'm
sure he and I will soon come back to that subject as we are going to be
working with him on integrating LVS into our product line. So after the
cacheoff next week and I actually have a few miutes to think about
something else, you'll begin to see some benchmark numbers on my
personal site as we work on that project.
http://polygraph.ircache.net
Note that the L4 switch workloads are not yet production quality, but if
you're willing to dig in and learn a little about workload creation,
Alex (primary author of polygraph) is very helpful.
Polygraph answers every need I can think of wrt benchmarking L4's and
web caches. You'll need to use a simulated proxy environment, probably,
to test an L4 switch. Pretend that your balancing between multiple
Polygraph proxies with the LVS and the BigIP and see which one is most
effective.
But beware that both BigIP and LVS will probably eat your test benches
alive long before you hit their throughput capability. Even a low-end
balancing box can handle far more packets than a couple of high end
benchmark client/server pairs can produce. That's a good thing...it
means that either will probably handle any sane load you can throw at
it...so multiple T3's or something is not out of reach for LVS, or BigIP
for that matter.
Note that Polygraph uses artificial real servers. Using real web
servers ruins any notion of predictability and makes it impossible to
isolate L4 peformance from server performance. Use a traditional web
server performance benchmark for testing the individual servers. If the
balancer is doing it's job you'll see linear scalability as you add all
of the realservers together (i.e. benchmark each one and then add the
results--that's what your cluster will handle). I don't know where to
point you for such a web server testing suite.
Good luck!
Jeremy Hansen wrote:
>
> Absolutely! This is very cool.
>
> Now...the position I'm in...I work for a fairly open company, but they
> still want numbers unfortunately and stats in comparison to other
> products, in my case BigIP's don't seem to be available.
>
> Even if the bigIP has slightly better performance, the cost difference is
> large enough that performance hits are tolerable, but I still need a way
> to come up with some number. BigIP's are like $15,000 a piece, 1U's
> running LVS is about $1200 a piece.
>
> The problem I've run into is the lack of free, good benchmarking
> utilities.
>
> What I envision is something that runs on multiple real client machines
> and not only get like http information, but also information from each
> one of the real servers and client machine.
>
> So, if you see performance drop, what is the reason, LVS bottleneck, real
> server bottleneck, network bandwidth bottleneck. I haven't found a good
> all in one tool to give this type of information.
>
> Even if I do win people over with lvs and we drop those bigIP, an
> important factor is the ability to determine how many real servers are
> needed for a given site or application. The company I work for turns out
> a lot of other companies and each one is unique in its application, so
> this is something that seems hard to determine.
>
> Maybe these aren't really completely lvs related issues, but I would like
> to use lvs and I need to proper ammunition to make it widely accepted.
>
> Thanks for any suggestions.
--
Joe Cooper <joe@xxxxxxxxxxxxx>
Affordable Web Caching Proxy Appliances
http://www.swelltech.com
|
