Building a robust server, what counts the most?

[JohnBee]

I am in the process of building a co-located 1u webserver. The target is to produce a server capable of hosting 25K users and 500 - 750 simultaneous logins. I am leaning towards a quad core Opteron system and I'm wondering if I should opt for 16GB of DDR400 ram or 32GB of DDR266? also, I wonder if there are clear cut benefits between SCSI over SATA(enterprise class) 10K drives in terms of performance?

And finally is there anything else I would need to consider?

[Zachery]

Instead of building one mosnter of a machine I'd build two, and co-locating tends to have a higher TCO than leasing does, esp when it comes to hardware going bad and trasitioning to new servers.

But yes, there are very very differnt preformance differences between SCSI and SATA, theres even a noticeable difference between SA-SCSI and SCSI.

[JohnBee]

Thanks for the info Zack, I did a bit if number crunching and under our current hosting a 3 year hardware investment yields significant cost advantages over equipment leasing. We are however planning to move to a full cabinet and add a second rack as our business grows. Thats the plan at least :)

So SCSI it is then, I wondered also if it was feasible to run virtual accounts under our primary root to protect or separate our accounts. I'm not totally up to speed of the CentOS resource allocation tools so this may be be nothing more than premature thoughts at this stage.

What about the ram? are we dealing with a more the merrier scenario or are there limits to the performance advantages of bulking up on cores and ram under the circumstances?

Instead of building one mosnter of a machine I'd build two, and co-locating tends to have a higher TCO than leasing does, esp when it comes to hardware going bad and trasitioning to new servers.

But yes, there are very very differnt preformance differences between SCSI and SATA, theres even a noticeable difference between SA-SCSI and SCSI.

[jason|xoxide]

I run my sites in virtual environments using OpenVZ. You might read a bit more about what I've done at my blog.

Stretching your vBulletin Servers with OpenVZ (http://www.jasonlitka.com/2006/12/13/stretching-your-vbulletin-servers-with-openvz/)

As to the RAM, it all depends on what the server is going to be doing. If you are going to run a 64-bit distro where your processes can actually access more than 2-3GB of memory then the 32GB may pay off, especially if you plan to preload your MySQL indexes into RAM. If 32-bit, or if you simple don't believe that you'll need that much RAM (32GB is a LOT, even if you do preload on a VERY big board), go with the faster stuff.

I have to ask though, why are you still using DDR? The newest Opteron and Xeon CPUs all use DDR2. Are you planning on going with older CPUs?

[JohnBee]

Hi thanks for the feedback.

Were headed for 64bit CentOS and so I guess more ram would be beneficial in our case then. To answer to your question, yes. DDR2 is the ram we would be using. I forgot to specify that. - OpenVZ sounds great! after reading on your blog a bit I wondered if such a setup(virtuozzo) would allow a user to dedicate cores or CPU affinity to one particular VPS? this could really enhance performance whereas it would take two separate machines to host db and front end data. Now with such setups we could assign two cores to one VPS and two other cores to another(discrete resource loads) without the cost and footprint of a dual rack setup. :)

Does this makes sense? or am I living in Technicolor again

[jason|xoxide]

Hi thanks for the feedback.

Were headed for 64bit CentOS and so I guess more ram would be beneficial in our case then. To answer to your question, yes. DDR2 is the ram we would be using. I forgot to specify that. - OpenVZ sounds great! after reading on your blog a bit I wondered if such a setup(virtuozzo) would allow a user to dedicate cores or CPU affinity to one particular VPS? this could really enhance performance whereas it would take two separate machines to host db and front end data. Now with such setups we could assign two cores to one VPS and two other cores to another(discrete resource loads) without the cost and footprint of a dual rack setup. :)

Does this makes sense? or am I living in Technicolor again

I do not believe that you can set a VE to use a specific CPU core. However, you can flag it to only be able to use a percentage of the total CPU power (meaning that 200% on a quad-core system would only give you access to half of the total CPU cycles).

[JohnBee]

I do not believe that you can set a VE to use a specific CPU core. However, you can flag it to only be able to use a percentage of the total CPU power (meaning that 200% on a quad-core system would only give you access to half of the total CPU cycles).
Your correct, after some more research we were lead to vmware esx. I looks like it will allow us to fine tune several independent OS's and get total control over memory and processing power etc.

[jason|xoxide]

Your correct, after some more research we were lead to vmware esx. I looks like it will allow us to fine tune several independent OS's and get total control over memory and processing power etc.

ESX does allow quite a bit of fine tuning, but it is VERY expensive and there is more overhead than with OS-level virtualization.