|
Hyper-Threading Performance Analysis - DivX and Conclusion
Hooz was gracious enough to dig out his copy of Antitrust so we could not only make comparisons between hyperthreading being enabled and disabled but also between 2GHz and 2.4GHz Xeons. Now this was the shocker I was looking for; disabling hyperthreading for our DivX encoding increased performance dramatically, especially in my testing with the slower Xeons. That 10fps difference between my 2GHz Xeons and my 2000+ Athlon MPs that I spoke of in my introduction has whithered away. In our DivX encoding, we frameserve from AVISynth into Virtualdub. While Virtualdub is not multithreaded, this method is an excellent way of taking advantage of multiple processors. I generally witness the Virtualdub process taking anywhere between 85% - 90% of my total processor time. Why does having hyperthreading enabled have a negative impact on our DivX encoding performance? That answer isn't known to me. I don't write code, and I'm certainly not an expert on operating system architecture (I'll leave that task to "jeh" in our forums). My suggestion to those of you using Xeon boxes for video encoding would be to do some of your own testing with whatever applications you use on a daily basis. You may note some surprises. Conclusion First, let me say that our exploration into the world of hyperthreading isn't complete. Today we've only really looked at benchmarks that affect the Windows workstation market. I'd like to delve into alternative operating system support (most notably Linux and FreeBSD) and server-oriented benchmarks. Anandtech previously surmised that more tangible performance benefits may be seen with hyperthreading in server applications such as database serving and web serving. I'd be interested to see if that holds true. What can we gather from today's testing? As we already knew, hyperthreading at present is a mixed bag with respect to performance. Certain applications will definitely see gains as we saw with Cinebench (based on Cinema 4D) and TMPEnc. We saw performance spiral in the opposite direction when encoding our Antitrust DivX with AVISynth and Virtualdub. In the end, you'll have to test the applications you use for yourself. The workstation market is all about stability and productivity. If disabling hyperthreading will increase productivity with the application you use at work or home, disable it. Having four CPUs displayed in task manager is all fine and good but certainly not reason enough to sacrifice performance which may impact your bottom line. End of story. That's all for today, but keep coming back because we're not finished examining Intel's hyperthreading here at 2CPU.com.
|