[vga]

Dear All,

For what its worth please find hereafter the results of my research about upgrading a Fluent configuration (8 licences pack).

We had a bi processor Intel Xeon X5650 @ 2.67GHz with 24,0GB Triple-Channel DDR3 @ 665MHz.

We replaced by a bi processor Intel E5-2643V2 3.5GHz 25M and populated with 8 bank of 8GB DDR3-1866 1Rx4 LP ECC REG.

The reduction in calculation time is ~42% (Hyperthreading MUST be disabled).

The processor was selected with 6 Core so with two processor with had 12 Core for a 8 licences pack. Those extra core are used for the system. The processor was also selected for its massive cache memory.

The memory was selected at DDR3-1866 and only 1Rx4 to ensure maximum speed. Also all the four bank of each processor were populated.

We are running Win7 Pro and fluent 14.5

We kept our 'old' graphical card and our two 1To HDD in RAID 0

The whole upgrade was below 4000eur.

Hope this help some people out here.

Regards,
Vincent

[ghost82]

Hi Vincent,
thank you for your report.
I have a question: you said that hyperthreading MUST be disabled; in my simulations on a bi xeon e5-2687w I always set the number of processors to 32 (so hyperthreading active, I can use 32 without license problem at university), because I though that there should be a minimum improvement in speed calculation.
Have you tried to compare times with hyperthreading on/off?

Thank you

Daniele

EDIT:
oh sorry, I understand that you can calculate on 8 cores "only".

[flotus1]

I would be more interested in a comparison on your Workstation between 32 "Cores"+HT on and and 16 Cores+HT off. Did you do anything in that direction?

[ghost82]

no, I never compared results with hyperthreading on and off, but I will do the test soon.
do you think hyperthreading must be disabled from bios to test it off?

[flotus1]

I dont know of any other method to deactivate Hyperthreading.

Edit: if you mean that you could leave HT switched on and compare the results simply by running on 16 cores and 32 "cores"... Personally, I would not trust these results.

[ghost82]

Ok, here are results from my small test, interesting results

Hardware:
2x Xeon e5-2687w (frequency 3,4 Ghz, turbo mode on on all cores)
64 gb ECC ram, clocked at 1600 Mhz

System: Windows 7 64 bit
Ansys version: 15 (all service packs applied)

Solver: stationary, pressure based, 2d planar, double precision
Velocity formulation: absolute
Models: standard k-epsilon, scalable wall functions
Phases: 1 phase
Boundary conditions: velocity inlet - pressure outlet
Scheme: simple
Discretization: gradient: least squares cell based; pressure, momentum, turbulent K.E., turbulent dissipation rate: second order
U.R.F.: defaults
Convergence at: scaled residuals <=10^-4

Total number of quadrilateral cells: 416000

TEST 1: Hyperthreading on, number of threads: 32
TEST 2: Hyperthreading on, number of threads: 16
TEST 3: Hyperthreading off, number of threads: 16

TEST 1
Iterations to reach convergence: 8727
Time to solution: 19 minutes and 30 seconds

TEST 2
Iterations to reach convergence: 7832
Time to solution: 19 minutes and 58 seconds

TEST 3
Iterations to reach convergence: 7832
Time to solution: 15 minutes and 1 second


ITERATIONS/SECOND
TEST 1: 7,5
TEST 2: 6,5
TEST 3: 8,7

N.B.:
Different number of iterations to reach convergence depends on different grid partitioning.
Parameter 'iterations/second' should be used to compare results.

So it seems hyperthreading OFF is better then turning it ON.

Daniele

[flotus1]

So it seems hyperthreading OFF is better then turning it ON.

That is exactly what I suspected. Thanks for your effort.

[ghost82]

If the time needed for one iteration is >> 1 sec maybe hyperthreading could be better, since I read some cases where it's stated ht on is better...I will try some cases when I will have some time..

[evcelica]

Originally Posted by ghost82

If the time needed for one iteration is >> 1 sec maybe hyperthreading could be better, since I read some cases where it's stated ht on is better...I will try some cases when I will have some time..

Turn hyperthreading off. If you have licences for 32 cores. add more nodes, otherwise you are just wasting licences trying to hyperthread. Use 32 real cores if you are serious, not 16 real and 16 virtual, which gives you no added memory bandwidth.

[RobertB]

We had a similar issue with opensuse and CCM+.

The issue seems to be that the OS did not seem to recognize when it was sending two threads to the same physical core and leaving another physical core open.

In general there seems no good reason to leave HT on, as you would never want to overload the physical cores for CFD.

If you did leave HT on you could set core affinity with MPI which should also solve the problem. Setting affinity on a non HT system didn't seem to speed things up even though you would think it might improve cache efficiency, etc.

[ghost82]

About Hyperthreading this is what is reported in the Ansys 16 preview installation guide (I don't know if this has been written also in previous version):

Hyperthreading technology uses one processor core to run more than one task at a time. ANSYS does not recommend using hyperthreading technology in conjunction with ANSYS CFD Solvers (Fluent, CFX and AIM Fluids). We recommend that you turn CPU hyperthreading off (default is on). A system administrator needs to reboot the system and enter the BIOS to turn the hyperthreading option off.