[DCN-TechOps] Base Specs for App Server

[Bill Broadley]

On Sun, Mar 14, 2004 at 04:06:11PM -0800, Steve McMahon wrote:
> Thanks for the suggestions!
> 
> We're already fronting Zope with Apache, and use it to cache what can be 
> cached. That helps with public pages, but still leaves us hurting on 

Using Rewrite rules?  ProxyPass?  PCGI/FastCGI?  CGI?

> items that can't be cached -- like management interface pages and 

But these should be very rare indeed right (when compared to user pages)?

> election result pages. (Also, I'm not seeing as much benefit from 

Why can't these be cached?  Even if the cache is set to 60 seconds
it could be a signficant performance increase.  

> The machine we're currently using as our main Zope server has four 
> 166mHz processors. I'm pretty sure that Zope/Python's threading isn't 
> making good use of the multiple processors (a known problem with 
> Solaris), so this is acting like a slow, single-processor machine as far 

Ah, Solaris.  Even if Solaris did do SMP well, having 4 processors
share a single 64 bit 66 MHz memory buss is going to lead to major
memory contention issues.

In comparison for instance the dual opteron should have around 48 times
more memory bandwidth.

> as Zope's concerned. One possible improvement might come from using ZEO 
> and binding the Zope client and the ZODB database to different 
> processors. But I think that would just be a holding action.

Yeah, I suspect a $400 dell server with a P4 and a 800 MHz FSB would be
substantially faster then a quad p6.

> Another consideration is that we'd like to be running some Plone sites. 
> I don't know if you've worked with Plone, but its ZPT pages typically 
> take about ten times as long to serve as DTML pages. (That also means 
> that they benefit even more from caching.)

I'm interested in Plone and am considering it for a new computational
portal for CSE.

> I hadn't thought about Opteron options, so thanks for pointing us in 
> that direction. The memory use benchmarks make it look like it might be 
> a great Zope platform.

I've seen web serving (both static and dynamic) as well as database benchmarks
that show substantial advantages as well.  I try to avoid over generalizing
microbenchmarks (like all the memory benchmarks), but certainly a killer
memory system doesn't hurt.  The dual opteron definitely has the
best memory system out currently, well except for a quad opteron.

I'm not sure where DCN sits as far as the whitebox hardware vs 
Tier-1 supported hardware.  Sun supports linux on the dual opteron today
and has announced support for dual opterons (the SunFire v20z) this
summer.

-- 
Bill Broadley
Computational Science and Engineering
UC Davis