I'll come right out and say this: I'm an AMD kinda guy. My main workstation runs Opteron 2220s, I prefer Opterons in my servers, and I've been looking forward to Barcelona for, well, far too long.
My attraction to the Opteron has been solely based on performance. In the past few years, the Opteron has consistently outperformed Intel's offerings, at least with my workloads. It seems that this may no longer be the case. My Stoakley reference system running two 45nm quad-core 3.0Ghz CPUs simply screams. I've had it running VMware ESX 3.0.2 for the past month, running a LAMP-based Web app load simulator I wrote in PHP. Even though on the face of it, Intel's design seems to be inferior and placing far too much emphasis on shared busses, the reality of the Stoakley platform is that it delivers.
Today's announcement from Intel regarding the new 45nm chips marks a reaffirmation of the next generation of processors. In reflecting on the fact that the new 45nm chips have transistor densities that place 30 million transistors on a space the size of a pinhead, I realize that the first Intel processors had around 2,300 transistors total. That was only thirty years ago, or so.
But enough reminiscing. Penryn is here, and it's going to make a splash if the performance of my reference system is any indication. The 6MB cache per die is definitely helping the numbers here, along with the 1.6Ghz FSB. If you want an exhaustive account of the ins-and-outs of the Stoakley platform and the new 45nm chips, check out Scott Wasson's Tech Report entry from September. I was at Intel in Oregon, standing a few feet from him when he took that beautiful photo of that wafer with a video camera of all things. I think I was laughing at the sight of three geeks desperately trying to photograph a largely reflective surface with digital cameras, and even took a photo of them trying to take photos.
On the outside of the box, I'm more concerned with what Stoakley can do for me. My Web app test hasn't changed much since I used it to test blade servers and VMware VI3 late last year. It's a relatively simple PHP script and a 500,000-record MySQL database. When a Web load generator is pointed at the script, it either serves up a static page, or makes a database call to generate a dynamic page. Tuning the parameters of the script can lean more towards Web server or SQL server performance, but normal operation has the predominate load shifting between the two as the test runs.
So I built a VMware ESX 3.0.2 server on the Stoakley box. The box itself had 16GB RAM, two 136GB U320 SCSI drives, and two Intel gigabit NICs. The onboard NICs weren't supported by VMware until just a week ago or so, and the SuperMicro case was low-profile, so I was limited to the dual-port LP Intel NIC I had. I dedicated one NIC to VM storage, and the other as the front end for all servers. I then built out six CentOS 5-based VMs: A two-CPU MySQL server, four single-CPU Apache servers, and a single-CPU load-balancer. Using LVS for load balancing, I nailed the box with HTTP requests for the Web test app installed on all four Apache servers. The load on all boxes grew substantially, delivering up to 4,000 requests per second, a number that's very dependent on the parameters of the test script, but certainly showed that the box was running on all eight cores, and running them hard. Over 200 million requests and nearly 15 hours later, the box showed no signs of distress, and each VM was still running like mad.
I kept the tests running over the next week, simulating a scenario that should never happen in real life, but occasionally does. After finally stopping the attack, I let the system quiesce, and then proceeded to load it up with more VMs for other tests I needed to do in the lab. I've wanted to run other benchmarks on the server, like compression and encoding tests, but that means that I have to power it down. I won't be able to realistically do that for another few days -- at this point, it's already indispensable.
So for now, I'm lacking raw numbers on standard tests, but I can tell you that my love affair with AMD is on shaky ground. I don't yet have a Barcelona system to run against Stoakley, and I really need to run those tests before I make any hasty moves. From what I've seen of Stoakley and Penryn this past month, my AMD honeymoon may be over.