SMP Provides A Mirror Into Your Server Increasing RAM memory will let your system cache data that otherwise would be read directly from disk. How much RAM do you need on the server? A good rule of thumb is: all you can get. Seriously, for a basic file and print server, you should start with 24 MB of RAM for Novell NetWare and 48 MB for Microsoft Windows NT. Add more RAM as needed for any applications you'll run. If you're building a database server, you want the ability to cache the entire database in memory. Just remember, if you decide to fully populate a server that can take "up to 4 GB of RAM," you'll buy $100,000 worth of memory for a $10,000 machine. Choosing the right RAID level also will have a significant impact on performance. Database applications like Oracle or SQL Server can be tuned to the Nth degree to improve sequential processing. Oracle even saves its writes an d performs them simultaneously, instead of doing several smaller writes. In this situation, RAID 1 would be better than the RAID 5 configuration you'd use for a file and print server. Also, remember to always use hardware rather than software RAID, so the operating system isn't tied up doing parity checks. All this adds up to one thing: It's essential to balance the system between CPU(s), disk and memory. If you think you have enough RAM and your drive array is configured correctly but you still need more muscle, then--and only then--is it time to add CPUs. Just remember, don't add what isn't constrained. You won't double performance by adding a second CPU. A lot will depend on your application, but don't be surprised if you only see a 60 percent to 80 percent improvement with the second CPU, and 10 percent to 15 percent with the fourth. Scott Mason is an analyst with a major energy corporation on the West Coast. He's a member of the Silicon Valley chapter of the NPA. He can be reached at smason@nwc .com.
	SMP Servers And Clustering Solutions You've got your symmetric multiprocessing (SMP) server up and running with redundant drives, power, fans, NICs and more. You're bulletproof, right? Wrong! If a processor fails, you're toast. And any data that was in that CPU's cache is gone. But don't lose sleep over it: There is a solution. Server clustering--long available in high-end RISC-based servers--is now becoming a hot-button item for Intel-based server manufacturers. The current crop of clustering products offers a 1+1 fail-over solution. Servers from Compaq Computer Corp. and Stratus Computer are examples of proprietary hardware-based fail-over clustering. There are also software solutions, such as those from Octopus Technologies or Vinca Corp. Fail-over clustering, in its current form, provides high availability, but does not enhance performance. Should Server A crash, its cli ents can log onto Server B and access the disks attached to both machines. Many manufacturers are waiting for Microsoft Corp. to release its WolfPack clustering software, but it will still only be a 1+1 fail-over solution in its first release. Plans call for WolfPack to eventually support an N+1 solution in which several servers will be clustered for high availability. Eventually, WolfPack will provide improved performance by allowing clustered servers to share a single application and divide tasks among multiple servers.

Updated April 24, 1997