An insufficient amount of RAM can cause your SQL server to continually read data from disk instead of cached memory.
This will impact on query performance in most cases this impact will be significant. Having an appropriate amount of memory will allow SQL Server to process your queries more efficiently: it can hold more data in its cache
CPU choice is easier ... The Opteron quad cores are cheap at present - you want as many cores as you can get for the buck (so a 2 socket system is probably your budget limit), even to sacrifice on clockspeed.
Look at putting at least 3 drives in it, in a suitable RAID configuration.
Disk Drives and Disk Space
When planning your SQL Server Hardware Requirements it is important to remember that database applications can extremely Input/Output intensive.
Most senior SQL Server DBA’s would recommend using fast SCSI drives and quality disk controllers (or if you have no budgetary restrictions a SAN environment).
If you want redundancy and fault tolerance you should consider using RAID. To quote from Microsoft "The different Levels of RAID provide varying levels of performance and fault tolerance. Hardware based RAIDS provides more performance that Operating System RAID".
Some of the characteristics of RAID can be seen in the table below:
RAID level Description Characteristics
RAID 0 Disk Striping/ No Parity Excellent performance – Negative fault tolerance
RAID 1 Disk Mirroring/duplexing Read and write performance is good and fault tolerance is excellent
RAID 5 Disk Striping with parity Read performance is excellent. Write performance is moderate - Fault tolerance is excellent
RAID 10 Disk Mirroring combined with disk striping (RAID 1 + 0 ) Excellent read performance Excellent write performance Fault tolerance is excellent BUT can be expensive due to the number of disk required.
That's the limit of my advice, as I poorly understand the Spec suite of server tests ... so my advice is a bit generic ... sorry.
Thanks, thats actually pretty useful info. I should mention though that the database will need to handle hundreds of thousands of clients accessing it, and it needs to be very fast. If it would give a large performance increase, i can increase the budget a couple thousand dollars (im a little flexible with this). Do databases rely on cores or clock speed more? And if they rely on cores, would the 12 core Opterons be a good choice? Secondly, how much RAM is needed for a database of about 10+ TBs of data? And finally, is RAID with SSD's worth it (price and perfomance wise) over RAID with HDDs?
I have 200+ severs and 2 mainframes (starting from) 10' from me. We have IBM's HP, Compaqs, and Dells.
All the Dells have needed service, mostly for raid cards issues and I hear that now you can only use Dell branded drives in them.... Drives drop out of raid often.
The IBMs, we have more of these than any, have had some issues. Some have had memory die, a few had hdd's go. a few have had fans go out. The service is excellent and the guys are usually here in hours. Parts are here in hours or overnighted when they are not stocked in driving distance. I am very impressed with their service.
The HP's and Compaq's have been reliable. I can recall only 1 power supply giving up the ghost, a DLT V tape drive, and a few hard drives. Service is slow, but better than Dells, from what I've seen but this may be whats contracted.
I dont know if other manufacturers require 'their' brand HDD's or not. I dont buy them - I'm just the guy that gets to call for service when they break.
Seems to me the best way to proceed would be to call or email hp/ibm whoever and ask for a quote. For a purchase of that size they surely have someone who can assess your needs and make a recommendation. I'm not sure if its just because I'm used to small load servers but something with over 100000 people seems to me you would have serious issues and need a MAJOR server system and possibly a SAN with fibrechannel. But I'm clueless so perhaps you should talk to the people selling it.
how many people would be using it simultaneously etc what type of data and such would probably have a large impact.