Terminal Server with 8 CPUs and 8GB RAM: How does it scale?

Those of you who read this site often know that I don’t usually just post links to other documents.

Those of you who read this site often know that I don’t usually just post links to other documents. (After all, that’s what Google is for, right?). However, every so often I come across a paper that is fantastic and deserves recognition.

Today I’d like to point out a paper by Dr. Bernhard Tritsch, the Chief Systems Architect for visionapp GbmH, a German server-based computing consulting company. The paper, called “The Big Iron Test” presents the methodologies and results of a study of Terminal Server on big (8 processor, 8GB of RAM) servers. (I've written on scalability testing in the past using Citrix's server test kit.)

Here’s the best part of the visionapp paer: While there are many papers floating around the Internet that show the results of such a test, this paper shows the detailed methods they used to conduct the test. It details everything and includes the scripts for how they created the 150 user accounts, how they threw all the user sessions at the server (with the full source of the scripts they used) and how they handled user profiles (again, with scripts).

This paper also details their testing methodology, including which applications they ran and which performance counters they watched.

The analysis section of the paper contains many graphs that show the exact results they obtained (both with HyperThreading enabled and disabled and on Windows 2000 and 2003 Server). In general, this is one of the best papers on performance tuning that I've seen in a long time.

Download this paper in PDF format from visionapp's site by clicking here.

Join the conversation


Send me notifications when other members comment.

Please create a username to comment.

Next, lets try PIII 866 processors, 6 gigs of RAM and Office 97!

The older 8xWay machines running W2k have shown to host more that 130 users.. Geez, I've been able to get 100+ users running office 2k and adobe reader on a 2xWay system..

And the author is shock on the # of context switches?! There are 16 processors! (8 physical, 8 logical) of COURSE there is going to be a LOT of context switches.
Sure it's using older hardware, but I like their process, and I think that's the real value of this paper. But you do make good points...
I think the results are strange ...
I´ve seen an average os 70 users per server using older PIII 1.26GHz (no HT), 4 GB RAM, and Windows 2000 Server, using a lot of applications, SAP Office, People Soft, etc etc ..

I wonder Windows Server 2003 with a newest Dual Processor(HT) machine, 4 GB, and ArmTech could get more than 100 users with decent performance. Or am I dreaming ???? I am talking about average desktop applications, Office, Internet Explorer, SAP, etc etc etc ... maybe more than 100 published Desktops !

What do you think ??
I agree with the basic concepts, but the methology needs improving. Why use W2k3 and only Office 2k? Why not Office XP or Ofice 2k3?

It would also be nice for these type of paper to give a cost analysis... An 8 way system is extremely expensive compared to 2x and 4x... Many company pass up 4xway box as they can be 3 to 4 times more expensive than a 2x way.
I think there must be another reason for the slow system. Such a system have to deal with 100 user without problems. Maybe they just havent modified the PTE or something like that. And I wonder why they didnt use Roboserver to simulate userload.
But I also think it is good to read the way they did it.
In my opinion, the load that is generated by their logon script remains undiscussed. If you are stress testing a system, it is extremely important not only to describe how your load is generated but also how this setup applies to reality.
It is highly unrealistic to launch three applications with given documents and then leave the session open.

This paper is certainly lacking some scientific procedures.
they use the /pae switch and there isn't any problem with 4 gb barriers

i also read the 4 gb memory paper there you give the advise to "not use" this /pae setting (4x more page-tables)

do you still advise to not use this setting ?

want to download
i want to read this papper