New PC: biggest bang for the buck?

[twade]

Juan,

Congratulations on you new machine.

Thanks!  I'm hoping I can improve its performance.  There's a lot of potential tweaks in the BIOS.

Any chance you install a Linux distribution?

At present, I do not have an extra drive; however, I do have VMWare Workstation so I could install Fedora 14 as a virtual machine.  The VM will result in some processing penalty, but hopefully it will give us a better idea as to what is causing the slowdown.  I will keep everyone posted.

Wade

[twade]

To all,

I updated the BIOS to the latest version and this helped a little.  I still have more tweaks to make, but here are the latest results:

New dual processor Xeon x5650 2.66 GHz:
   Benchmark_M74:              62.60 (33 percent improvement)
   Benchmark_M74_parallel:  11.12 (9 percent improvement)

Wade

[twade]

To all,

I installed Fedora 14 as a VM into WMWare Workstation.  The parallel processing took a big hit; however, the non-parallel results were a little better.  Here are the numbers:

New dual processor Xeon x5650 2.66 GHz:
   Benchmark_M74:              42.35
   Benchmark_M74_parallel:  25.64

I'm still looking into what might be causing the poor performance.  I'm hoping I can continue to find ways to improve the results.  Despite the results being on the high side, the machine is significantly faster than my old machine.

Wade

[georg.viehoever]

It is difficult to diagnose such things. For sure, this machine should outperform my laptop in the sequential benchmark, but it does not....
Since you already upgraded the BIOS:
- try Linux (outside of the VM - within the VM at cannot be be faster than native Windows). If you dont want to install it, you should be able to run a live DVD (you have more than sufficient RAM)
- are all RAM slots occupied? If not, all RAM may be connected to only one of the CPU sockets, giving you bad performance for the second socket.
If you like, you can swap your dream machine against my old laptop 

Georg

[twade]

Georg,

It is difficult to diagnose such things.

Definitely.  Later today, I'll see if there are any firmware updates to the components on the motherboard.  This may help disk throughput.

you should be able to run a live DVD

I'll give the live DVD a try tonight and post the results.

are all RAM slots occupied?

No, only slots 1, 3, and 5 (i.e three channel mode).

If you like, you can swap your dream machine against my old laptop

What's your mailing address. 

I appreciate all your suggestions.  I'm positive we'll get to the bottom of this.  In all likelihood it is Windows.

Wade

[georg.viehoever]

Hi Wade,

I just had a quick look at the manual of your board:
- do you have DIMMs in both banks of the board? If not, then one socket would always need to fetch memory by asking the second socket, which is extremely inefficient (and from what I read not really supported by this board).
- your board has jumpers to disable CPU sockets. Maybe you check what happens if you disable one of the sockets.

Good luck,
Georg

[twade]

Georg,

do you have DIMMs in both banks of the board?

Yes, I have a total of six 4GB DIMMS.  They are located in slots 1, 3, and 5 on each bank.

your board has jumpers to disable CPU sockets. Maybe you check what happens if you disable one of the sockets.

I may have to go down this road if I all other leads fail to deliver what is expected.

I really appreciate your thoughts and suggestions so keep them coming.

Wade

[twade]

To all,

I ran Fedora 14 Live KDE and definitely got some improvement, especially with the parallel processing benchmark.  Here are the numbers:

New dual processor Xeon x5650 2.66 GHz:
   Benchmark_M74:              29.82
   Benchmark_M74_parallel:  6.70

Obviously, Linux makes use of the processors better than Windows; however, I'm not sure why my non-parallel benchmark is so low compared to others.  I did notice the following in System Info under Fedora:

Processor: X5650 @ 2.67GHz
Speed: 1,600 MHz
Cores: 24

What is the Speed measurement?  It seems way low.

Wade

[Andres.Pozo]

Hi, I have just run the benchmarks in my new machine:

Intel Core i7 2600K @ 3.4GHz
8GB RAM
Win 7 64bits
   Benchmark_M74:              20.14s
   Benchmark_M74_parallel:  7.176s

[georg.viehoever]

Processor: X5650 @ 2.67GHz
Speed: 1,600 MHz
Cores: 24

Wade,
this speed may be caused by the fact that idle processors are throttled. Look at the CPU speed while running some program, it should show the nominal speed or sometimes something faster (TurboBoost).

Georg

[twade]

To all,

I made several more BIOS tweaks and changed the power savings to High Performance in Windows 7.  This gave me a little more improvement.  I'm not sure why but this appears to be the best this system can do.  Even under Linux, the non-parallel numbers are high.  It's almost as though PixInsight isn't taking advantage of the Xeon 5600 series.  It will be interesting if I see any improvement with the next version of PixInsight.  I will keep you posted.  Despite the "disappointing" numbers, I'm still quite pleased since this system is significantly faster than my old system.

New dual processor Xeon x5650 2.66 GHz:
   Benchmark_M74:             55.12 (13.5 percent improvement)
   Benchmark_M74_parallel:  10.42 (6.7 percent improvement)

Wade

[twade]

To all,

As one last experiment, I turned off hyper-threading.  This improved the results in PixInsight, but there was a decrease in performance in other benchmark tests (e.g. CINEBENCH 11.5) which is quite interesting.  I wonder if I have reached a thread maximum for PixInsight.  Any thoughts?

New dual processor Xeon x5650 2.66 GHz:
   Benchmark_M74:             38.18
   Benchmark_M74_parallel:  9.22

The CINEBENCH 11.5 results are as follows:

With HT:
  CPU 13.97
Without HT
  CPU 11.93

With HT:
  OpenGL 39.79
Without HT:
  OpenGL 39.15

It appears there is a penalty for decreasing the number of threads for the CPU benchmark in CINEBENCH 11.5.

Wade

[Juan Conejero]

Hi Wade,

I wonder if I have reached a thread maximum for PixInsight

There's no specific limit; PixInsight will use all logical processors allowed by global preference settings (all processors available are used by default).

There's something that you may try in PixInsight:

- Select Edit > Global Preferences.

- On the Preferences tool, select the Parallel Processing and Threads section.

- Disable the Enable Thread CPU Affinity option.

- Press F6 or click the Apply Global button.

Thread CPU affinity control normally improves performance by preventing expensive cache invalidations caused by running threads that migrate across processors. With affinity control enabled, PixInsight will force the OS to execute each thread on a specific logical processor. This is a small performance improvement though, and it is OS- and architecture-dependent. I suspect this feature may be penalizing performance with these Xeon processors under Windows 7 for some reason.

Another, much more important performance penalty may come from the fact that the benchmark image is quite small (about 900x900 pixels). Most image processing algorithms are embarrassingly parallel. In most cases, with 24 logical processors each thread will process less than 40 pixel rows. This is a marginal load that may be causing a severe performance penalty because the work required to set up and launch the threads may be comparable to the actual pixel processing work. So my recommendation is:

- Re-enable hyperthreading in your BIOS settings.

- As before, open the Preferences tool and select the Parallel Processing and Threads section.

- Disable the Allow using all available processors option.

- Decrease the Maximum number of processors used value to 12 (for example).

- Apply Global.

Let's see if these changes improve benchmark performance.

Note however that your machine performs much better under Linux (with PixInsight at least) than on Windows 7. And if you fine tune services and background tasks on Linux, you can get significant performance improvements (the Live CD is far from running optimized). With a powerful workstation like yours I definitely would use Linux for all serious tasks and would install Windows, if necessary, in a VMware virtual machine. This is just what we do for PixInsight development. Another option is a dual-boot setup, but VMware is much more practical.

[georg.viehoever]

...As one last experiment, I turned off hyper-threading.  This improved the results in PixInsight, but there was a decrease in performance in other benchmark tests (e.g. CINEBENCH 11.5) which is quite interesting.  ...

Hi Wade,

I know quite a couple of applications where Hyperthreading does not give you any additional performance, and a few where it actually hurts. Maybe PI is one of these. So nothing unusual here.

Two more hints:
- You may want to monitor the actual CPU clock (and tempeature) while running the benchmark. Maybe they are throttled because of thermal issues? Something like Rightmark CPU Clock http://cpu.rightmark.org/ may help with that.
- Do you use ECC memory? If yes, try to switch off ECC.

Georg

[twade]

Juan,

I appreciate your suggestions.  I'll give them a try tonight.

Georg,

Maybe they are throttled because of thermal issues?

I've watched them using ELEET which comes with the motherboard.  The cores are very cool during the benchmark phase.  None of them get warmer than 25C.  The air temperature coming out of the machine is a cool 20C.  I built it just in case I OC it  in the future.  It has 15 fan including the power supply.  I'll post a few images of this beast later if anybody is interested. I'll check out the program you mentioned too.

Do you use ECC memory? If yes, try to switch off ECC.

Yes, I do use ECC memory.  I'll see if there is a way to run this feature off.

I appreciate all your suggestions.  We will definitely get to the bottom of this.  Is is obvious, Linux is a much better OS when it comes to speed.

Wade