[lawrencelaw]

Hi guys,

My current set up is an i7 920 Overclocked to 4Ghz with 6G DDR3 Ram and is mainly used for running FDS. I am looking into getting a workstation and for budget reasons, my 1st choice is an Skulltrail dual QX9775 with 16G ECC DDR2 Ram.

Any feedback will be helpful, my worry is the skulltrail will be a dud and i have received conflicting reviews on the web.

Is it true that by configuring the hardware, we can "make" the CFD softwares to populate all the 16 memory bandwidths (8 core with DDR2, 2 channels for each core, so that makes 16)? This is what i was informed.

Thanks
Best Regards
Lawrence

[Marcelo]

skulltrail vs i920 is very similar, with a i7 processor you have 8 thread and with a skulltrail you have the same, if you wanna improve the system the first choice is have 6x2gb ram ddr3... if you would like more ram or best cpu go for a i940 or better... skulltrail is like a old xeon, the new nahalem xeon are like the i7 (8 thread)

[Amiga500]

Forget about skull trail.

The Clovertown/Gainestown series bottleneck badly due to the front serial bus.


In all likelyhood, your i7 running on 4 threads will be quicker than the Skulltrail on 8!!

[lawrencelaw]

Hi,

Thanks for the feedback. I have purchased a dual 5580 system and have run the same CFD program.

You are correct, the i7-920@3.8Ghz is faster (5-10% approximately) than the dual 5580 but the comparison is still not conclusive due to the following factors:-

1. The dual W5580 has only 6G memory shared between the 2 cpu.
2. The CFD software can only populate 4 cores, maybe 5. So the W5580 is running only at 60% while the i7 is at 100%.

The motherboard i have does not allow me to OC the W5580, else it will be at 4Ghz for sure.

Regardless...I am happy with the dual 5580 as it is running about the same speed as the i7-920 at 3.8Ghz and with a new mother board (it will be next year's budget for sure!) that can allow me to OC, for sure it will be better than the i7.

Thanks for the feed back. By the way what is this bottle neck you were talking about...any sites for me to read more about this?

Thanks
Lawrence

P/s- If i had the foresight, I think you are right- 2 x i7 with cross over cables for parallel computing would be better i guess (I have yet the opportunity to determine how much better) than a dual W5580 in-terms of value for money unless the software can populate all 8 cores at once.

[Amiga500]

Originally Posted by lawrencelaw

2. The CFD software can only populate 4 cores, maybe 5. So the W5580 is running only at 60% while the i7 is at 100%.

OK - if this is the case, you should be sure to set your processes to particular processors.

You can see as much as a 15% speed-up by avoiding cache clearances due to the operating system screwing around and by maximising the cache you have available.

You've 8 cpus, these are labelled cpu0 to cpu7 within the operating system.

For 2 thread jobs, use:

cpu0 -OR- cpu1 -OR- cpu2 -OR- cpu3
-AND-
cpu4 -OR- cpu5 -OR- cpu6 -OR- cpu7


For 4 thread job use:

cpu0 -OR- cpu1
-AND-
cpu2 -OR- cpu3
-AND-
cpu4 -OR- cpu5
-AND-
cpu6 -OR- cpu7


Basically, you try to ensure each processor has the maximum possible cache available, and you also avoid needless cache clearances.


Try it back-to-back if you want to get a feel for how much difference it makes

[lawrencelaw]

Dear Amiga500,

In FDS software the more meshes i create, the software can allocate a core/thread resource to each mesh, That is if I have 8 meshes, in theory i can populate all the 8 cores at once.

I am not aware how to execute your advice on the cache re-allocation, could you provide or direct me to some guide that i can adopt? Much appreciated.

Thank You
Lawrence

[Amiga500]

You cannot reallocate the cache - but you can allocate your processes to certain cores to maximise your use of the available cache.


In windows, go to:

task manager >> Process >> <right click on a process> >> set affinity


Then follow the logic in the last post of mine.

[lawrencelaw]

Dear Amiga500,

Thank you for the input, the task manager instructions was clear but I am still unsure how to use the instructions you posted earlier. I will try to give an example here to see if I understood them correctly:-

I have 5 processes threads running now (because of the 5 meshes in the CFD model) with 4 cores on my i7 and 8 cores on the dual xeon.

When i right click on the 1st thread (in task manager) i can see the 4 core (core 0 to 3 for the i7) or 8 core (core 0 to 7 for the dual xeon). Based on your instructions, for thread 1 set to core 0, thread 2 to core 1, thread 3 to core 2 etc...?

Sorry i am still new to this..if my questions sounds ignorant it is because I am.

Thank You
Lawrence

P/s- thanks for the good advice, i think I am making a breakthrough here!

[Amiga500]

Originally Posted by lawrencelaw

I have 5 processes threads running now (because of the 5 meshes in the CFD model) with 4 cores on my i7 and 8 cores on the dual xeon.

Sorry, maybe I am mis-understanding your terms.


By 5 processes you mean 5 mesh - but each mesh can have multiple threads? (i.e. you are already using all 8 cores on the dual xeon)


What I described earlier only works if:

(1) You are running 1 job, and using 4 cores or less.
(2) Running 7 or less threads, but one job can reside on one core.


This is the cache hierarchy of one of your CPUs (c.f. Barcelona & K8), you want to maximise the L2 cache available to any core in use.

[lawrencelaw]

Dear Amiga500,

This is getting pretty technical....i guess i need to understand the diagrams you sent to get a clear picture of your explanation.

Moving ahead, when i ran the CFD job (with 5 meshes), the CFD program can potentially populate 5 cores i.e. it appears as 5 separate processes in the task manager, using the dual xeon as the example (HT is turned off so i get only 8 physical cores), by setting the affinity, matching one dedicated core to each of the 5 process, i can get each core (in-total 5 cores) to
run at 100% each all the time. Before this, the cpu loading was distributed and fluctuating among all the 8 cores.

Will what I did above improve the cpu performance as i "solved" the cache clearing problem and maximized L2 cache?

I hope to understand the 3 diagrams you provided to improve my understanding how the cpu functions.

Thank you for the valuable input

Lawrence

[lawrencelaw]

Hi,

Using the same CFD model (about 1.5milliion cells and Total CPU RAM utilization is less than 5G for both set ups) and with the following configuration:-

1. Dual W5580 at Turbo 3.3Ghz and 12G ram.
2. i7 920 OC 3.8Ghz and 6G ram.

For both the set up the CFD model is divided into number of mesh zones so as to populate all the available CPU cores (8 cores for dual W5580 and 4 cores for i7).

The W5580 takes about 64 hours to complete the CFD run (projected as it is still running)
The i7 920 takes about 106 hours to complete the CFD run.

The dual W5580 is about 65% faster than the i7.

I would think if the W5580 can be OC to 4ghz, it would be about or slightly more than double the i7 performance.

W5580 temperatures is running at about 62 to 72 degC now, avg about 70degC.

Hope this info helps for those who are considering getting a dual W5580.
I guess the next step is that to compare two i7 OC 3.8Ghz linked by lan cable vs a single 2xW5580 performance.

Lawrence

[cpu]

Originally Posted by lawrencelaw

Hi,


1. Dual W5580 at Turbo 3.3Ghz and 12G ram.
2. i7 920 OC 3.8Ghz and 6G ram.

For both the set up the CFD model is divided into number of mesh zones so as to populate all the available CPU cores (8 cores for dual W5580 and 4 cores for i7).

The dual W5580 is about 65% faster than the i7.

Lawrence

Dear Lawrence,

Thanks for your post. Right now I'm thinking what to buy. Options:
1. Dual W5590 6x2Gb DDR3-1333
or
2. i7-975

You noticed that you run 4cores on i7 - it means with hyperthreading off?
I understand that with hyperthreading 8 processes are availiable. Could it be faster?

Also, I'm not sure with memory configuration for both. Intel talksabout tripple memory channel - so should I use six memory sticks? (3 channels x 2 W5590 sockets)

Thanks,

Vladimir

[bramv101]

Hyperthreading is not always available for CFD software. Either way its not the way to go.

Go for a dedicated machine in stead of a gaming machine

[cpu]

Originally Posted by bramv101

Hyperthreading is not always available for CFD software. Either way its not the way to go.

Go for a dedicated machine in stead of a gaming machine

Thanks for the comment. I know that hyperthyreading is controversal. Threads will fight for the cash memory, etc. However, if the program algorithm can be well divided in parallel threads - we can benefit.

xeon CPUs are dedicated mainly for data/transaction servers -they are very reliable, yes, but expensive. You have to use expensive RAM too. Yes, i7 is advertisied for "gaming" machines but it also fast in FP operations. If you only need high-speed calculation - it can be a choice, IMO.

I write the code myself - so I'd like to try hyperthyreading and compare i7 vs xeon. if i7 is faster or the same speed as double xeon - why not to use i7? I don't need 100% ready, mirrored RAIDs neither hot drive exchange nor automatic backup.

My questions are
- has anybody tried to switch hyperthreading on for xeon?
- how smooth the change, OS, etc.
- does you CFD code recognize 16 CPUs then?
- does it make a diffrence in calculation time? - not CPU time - but real time?

Thanks,
Vladimir

[lawrencelaw]

Hi,

I have tried hyperthreading with my i7-920. I turned it off for the following reasons after a month of use:-

1. The core temperatures were a good 10degC lower with HT turned off.
2. The HT feature did not help to reducer the CFD computing time, i had a feeling it was slower with HT on- but no hard figures to back this up.

Lawrence

Ps- i experimented running CFD with the dual W5580 and it was slower with HT....could not be sure about the exact figures as i did not note them down but it was definitely a decision maker for me to turn HT off.

[cpu]

Originally Posted by lawrencelaw

Hi,

I have tried hyperthreading with my i7-920. I turned it off for the following reasons after a month of use:-

1. The core temperatures were a good 10degC lower with HT turned off.
2. The HT feature did not help to reducer the CFD computing time, i had a feeling it was slower with HT on- but no hard figures to back this up.

Lawrence

Ps- i experimented running CFD with the dual W5580 and it was slower with HT....could not be sure about the exact figures as i did not note them down but it was definitely a decision maker for me to turn HT off.

Thanks. Dreams never come true.

It is known that for a single thread HT off is better, of course.

Do you remember if your CFD code (with HT on) recognised 16 CPUs on dual Xeon and 8 CPUS on i7? I hoped 16(8) parallel threads withy HT could be faster then 8(4).

Vladimir