Unix/Linux Pre-Install Questions to Ask and Answer

• What is the purpose of your system?
• Will a single system be enough (or SMP or cluster)?
• Do you need to consider virtualization (such as Vmware or Xen)?
• Should you spend the money on SCSI, RAM, big disks, ...?
• Should you use SANs, NAS, an internal RAID, or what?
• Will you be building a network (such as SOHO or larger)?
• What kind of network to get (if any)?
• Where will you put the equipment (a server closet a.k.a. data center)?
• What kind of racks to get?  Wire supports?
• What other equipment is needed (monitor switches, carts, testers and monitors, ...), UPS, etc.
• Do you have the correct physical systems (facilities) in place or will you need to upgrade or add HVAC, fire suppression, security, etc.?

Capacity planning:

Some vendors will tell you to buy a large number of servers, run the system awhile to see how it performs, and return the unneeded servers for some sort of refund.  But there are other ways to estimate accurately the number of servers required for some service(s) to run and provide a required average response time (sometimes other performance requirements too).  These include:

• Build a prototype system, measure performance and response time, and extrapolate the required number.
• Trust the vendor to only sell you what you actually need.
• Estimate the resources needed for each part of the system (e.g., web, forms, email, etc.) and sum these values to obtain an estimate of the whole system requirements.
• Calculate the value needed based on the factors of the servers you are considering: number, type, and speed of CPUs per server, amount and speed of RAM and disk system, network bandwidth available, and the system's I/O performance data.
• Hire an experienced consultant who does this sort of planning.  Speak with the former customers (from a year ago) to see that they were pleased with the results.

Without experience, the required mathematical skills, or a huge budget to blow on consultants or needless equipment, you can use queuing theory to accurately estimate the number of servers required to provide some specific response time, by plugging in a few values into a fomula and graphing the result.  There is a formula from Dick Brodine of National Computing Group (DickBrodine@juno.com), reported in the July 2006 IEEE Computer (“Mathematical Server Sizing”, pp. 91–93).  This formula is based on one by Kishor S. Trivedi in his book Probability and Statistics with Reliability, Queuing, and Computer Science Applications, ©1982 Prentice-Hall.  Trivedi's model was designed to estimate response time from a number of factors, assuming a single server.