Hewlett-Packard's DL 585 Server
HP introduces its DL 585 server with four 64-bit AMD Opteron 880 dual-core processors.
March 10, 2006
The multicore rush is on, and what a race it's going to be if the performance of Hewlett-Packard's DL 585 server with four 64-bit AMD Opteron 880 dual-core processors is any indication.
I installed the beefy 4U server in our Green Bay, Wis., Real-World Labs® for testing. The 2.4-GHz CPUs run 64-bit Windows Server 2003 on 16 GB of RAM and 138 GB of SCSI disk managed by HP's Disk Array management system, 4-Gbps NICs, and a PCI-X bus. The server has six free PCI-X slots--expansion space that's increasingly lacking in servers--for peripheral cards. You could quibble with the lack of PCI-Express slots; HP based the server on its year-old single-core machine, PCI-X bus and all. HP wanted to be among the first to market, as did competitors such as Sun Microsystems, whose 800-series dual-core Opteron, the SunFire V402, also uses a PCI-X bus.
Starter Tests
With the server's memory, fast disk, wide network bandwidth, and four dual-core CPUs, I knew testing it wouldn't be as simple as just plugging in some gear to learn where the machine is overloaded. This is the first time Network Computing has had hands-on experience with a dual-core Opteron system, but all servers have enough in common that I knew my first test would be to point Spirent Communications' Avalanche 2500 load-testing appliance at the DL 585 and, using Microsoft's IIS, hammer the machine with a 24-KB test consisting of an HTML file and four images. Given IIS' ingrained inability to put a load on more than four processors, I didn't expect this test to make much of a dent on the server, and it didn't.
Good • Outrageous CPU performance from the dual-core Opteron 880s • Four teamable 1-Gbps NICs• Highly expandable memory Bad• CPU performance may be too much for your needs • PCI-X slots only, no PCIe• Dated internal disk subsystem may hold back performance DL 585 Server, starts at $8,387 ($31,667, as tested). Hewlett-Packard, (650) 857-1501. www.hp.com |
DL 585 Testing Click to enlarge in another window |
This server has so much processing power that the CPUs' maximum usage registered at just 54 percent, and averaged only 10 percent, while the network sat pegged at 100 percent throughout the test. Not to be daunted, I set up a test to maximize CPU usage. I sent 1,082-byte files, and the system again maxed out the network. I sustained an average of 30 percent overall CPU usage. Nice.
I used several other test tools, trying to find something that would give the DL 585 a run for its money. Unfortunately, test tools are lagging behind CPU advances, and those available to me didn't adequately stress all eight CPUs, didn't handle 64-bit CPUs, or couldn't generate enough work to overload the CPUs. On the bright side, I now have a new item on my list of "must have" tools--a 64 bit, multi-core compatible test tool for servers.
Strength To Waste?
As it was, I still learned something valuable from these tests. Unless you have a processing-intensive application, or you're interested in consolidating servers using virtualization software, this system is likely to be wasted. Throw one of these machines up as a garden-variety Web server, and it will sit with CPU time idle more often than not, even if you flood the network pipes. With a database server, the disk, not the CPU, is the limitation. I ran 32-bit IOMeter tests that only blipped CPU usage, hitting a maximum of 6 percent utilization, with a 3 percent average, while maxing out disk throughput.As a virtual server, though, this machine would rock. Virtual servers use CPU time more than most applications, and require more memory than other servers generally do, and the DL 585 has plenty of CPU power and RAM capacity.
Throughout testing I watched disk accesses so that I was certain where the bottlenecks were in the system, and the test system's single disk appeared to be the bottleneck in at least one of my Spirent-based tests. Disk usage began to spike relatively early in the run cycle of the 1-KB file test, and continuously spiked to 90 percent before dropping back to near zero. This is generally a sign of a disk being stressed in a multitasking operating system. Since this was a server and CPU test, I didn't hook the server up to our Fibre Channel SAN and run tests against it. This means that the disk was running Ultra 320 SCSI on a small number of spindles, so it's no wonder disk became the bottleneck early in the game.
Potential buyers should remember that there are few applications out there that will fully utilize dual-core systems, but don't let that stop you if you're in need of a powerhouse machine that can handle the CPU load of about anything you throw at it.
Don MacVittie is a senior technology editor at Network Computing. Write to him at [email protected].
You May Also Like