High-powered 1U servers make the grade

Many administrators assume that 1U servers are intended only as nodes in a Web farm or high-performance computing cluster, where high availability is a function of the cluster rather than the individual node. However, a number of 1U servers are now available that include features most admins want in application or file servers, including redundant power supplies and fans, RAID controllers, and dual processors.

I looked at four such systems -- from Gateway, Hewlett-Packard, MPC, and Supermicro -- to examine the management, redundancy, performance, and expandability they offer, and I was pleasantly surprised. (Other vendors with similar systems, including Dell and IBM, were unable to participate.)

Testing: One, two, three ... four

I received the Gateway 9415, HP ProLiant DL360 G4p, MPC NetFrame 1720, and Supermicro SuperServer 6014P-TR for my test. All of the servers support 2.8GHz to 3.6GHz single or dual Xeon processors, have at least 12GB of RAM, and come with dual redundant power supplies.

Expandability is fairly limited because of the servers' 1U size -- one or two PCI-X slots, three or four drives max. All but the HP included an external SCSI port for adding direct attached storage.

These small servers offer a couple of surprises: First, the SATA-based MPC and Supermicro servers offered file server performance as good as or better than the Ultra320 SCSI systems; second, the most capable system, the Supermicro, was also the least expensive.

I set up the systems with Windows 2003 Standard Server and created a file share and then copied folders with either numerous small files or a few very large files, using several users on different client computers to copy files to and from the file servers. This test simulates conditions that real file servers face more closely than tests run locally on the file server using Intel's IOmeter, which doesn't necessarily reflect real-world file server performance.

The results for the file-server tests were interesting. In throughput tests using large files, the Supermicro system had the highest throughput, averaging just less than 65MBps over a gigabit link. The Gateway was second, at 57MBps, and the HP and MPC tied at almost exactly 42MBps.

However, it seems that the number of drives had a greater impact on the servers' performance than the drive interface did. In tests using many small files, I copied approximately 30,000 files using eight clients simultaneously and measured the time to complete copying all the files. Here the Gateway came in first, with a time of 37 minutes. The Supermicro was second, with a time of 41 minutes; the HP finished in 48 minutes; and the MPC in 52 minutes.

As for what caused this big difference in performance, the basic limitation in a small system is the performance of individual drives -- the more drives you spread I/O over, the better the performance will be. The disk itself is the bottleneck, not the controller or drive interface. As this test shows, that is true both in terms of overall throughput and I/Os per second. One might expect the Ultra320 systems to do better in I/Os per second, which they usually do in IOmeter tests; this might have been the case if the Ultra320 systems had had similar numbers of drives, but with fewer drives, they couldn't keep up.

Comparing and contrasting