321 LLR timings

[paulunderwood]

Here are my current 321 LLR timings around 2.6M bits:

FFT-length     CPU              MHz       Hours/LLR test 
----------------------------------------------------------------

131072         PIII             980      15:00
131072         Athlon          1050      10:10
131072         AthlonXP        1400       7:20
131072         AthlonXP        1666       6:30
131072         AthlonXP        1800       6:15
131072         AthlonXP        2200       5:00

163840         Pentium4        2500       4:30
163840         Pentium4        2940       3:50
163840         Pentium4        3500       2:45

[Thomas11]

Here is some addition to Paul's list (all at n=2.6M):

FFT-length     CPU              MHz       Hours/LLR test
----------------------------------------------------------------
131072         PII              300      50:00
131072         Athlon          1000      11:20
131072         AthlonXP        2000       5:40
                                                                                
163840         Pentium4        1700       6:35
163840         Opteron 246     2000       5:00
163840         P4/Xeon         2400       4:20
163840         P4/Xeon         2800       3:45
163840         Pentium4        3200       2:55

[paulunderwood]

With the new LLR 3.7 I have noticed a significant speed up on some pentium4s (but not much on my 3.5GHz):

MHz        3.6.2          3.7
-------------------------------
2500       4:30          3:50
2940       3:50          3:35

Thanks Jean

[paulunderwood]

I have egg on my face because it seems the timings above refer to LLR 3.6 and not LLR 3.6.2. If I had upgrade earlier I would have saved loads of time...

[paulunderwood]

I now have some timings for LLR tests for 3*2^n-1 at 1 million decimal digits on an Asrock 4coreDual-Vista mainboard.

FFT-length   CPU          MHz       RAM             LLRs    mSecs/Iteration 
-----------------------------------------------------------------------------------------

196608       PentiumD     2666      666Mhz Single    1      6.407 
196608       PentiumD     2666      666Mhz Single    2      6.800

196608       Core2 Quad   2400      666Mhz Single    1      4.017
196608       Core2 Quad   2400      666Mhz Single    2      4.022
196608       Core2 Quad   2400      666Mhz Single    3      4.098
196608       Core2 Quad   2400      666Mhz Single    4      4.168

196608       Core2 Quad   2400      666Mhz Dual      1      4.021
196608       Core2 Quad   2400      666Mhz Dual      2      4.017
196608       Core2 Quad   2400      666Mhz Dual      3      4.093
196608       Core2 Quad   2400      666Mhz Dual      4      4.153

196608       Core2 Quad   2400      533Mhz Dual      1      4.015
196608       Core2 Quad   2400      533Mhz Dual      2      4.017
196608       Core2 Quad   2400      533Mhz Dual      3      4.163
196608       Core2 Quad   2400      533Mhz Dual      4      4.172

The core2 quad does better at handling more LLRs than the PentiumD. That is running more than 1 LLR does not drastically change interation times.

Running dual channel as opposed a single stick of RAM makes little difference.

Memory speed (on this board) makes very little difference.

Clock for clock, using all cores, the core2 quad is 80% quicker per core than the PentiumD.

The quad is older type using 105 watt; whereas the Pentium is 95 watt. The latest 2.4GHz quads are 95 watt.

One of my current test range results shows that the PentiumD was crunching less than 7.6 "321" numbers per day. The core2 quad does 23.8 numbers per day. This means the quad has over 3 times the throughput

[Cruelty]

Have you tried running LLR on a single core? Could you post some results of your configurations running only one instance of LLR ?

[paulunderwood]

Note the "#LLR" column. This is the number of LLRs being run. (North Americans use "#" for number.) If #LLR is "1" then 1 LLR was running and, apart from a processes from a minimal Debian with no X, no other number crunchers were running.

Edit: I might have used the "#" improperly and so I have changed the column heading to "LLRs"

[Cruelty]

OK, if "LLRs" means how many instances are running at a time then performance hit is really negligible, which is a little bit odd looking at the specs of memory subsystem.
What are the system's specs in each scenario, I mean motherboard + memory timings?
OK, I'd like to run this "benchamark" on my C2Q system. Could you post a fragment of your input file, so that we can make an apple-apple comparison?

[paulunderwood]

What are the system's specs in each scenario, I mean motherboard + memory timings?

The board is running at default FSB but with things like on-board sound switched off as this is a dedicated number cruncher (running the OS from a USB stick.) The 256Mb memory sticks are 667 DDR2 from HP I think -- nothing special. It was set to SPD -- 5-5-5-15? Look at the "RAM" column to see where I ran it in either dual channel or with a single stick, and where I dropped it down to 533MHz.

OK, I'd like to run this "benchamark" on my C2Q system. Could you post a fragment of your input file, so that we can make an apple-apple comparison?

Yes please! The two-line file is attached.

Ps. the iteration times given are the averages for the various number of LLRs running.

[Cruelty]

OK, here are the timings for my system C2Q @3186, RAM@1065:

• 1xLLR = 2.859
• 2xLLR = 2.922
• 3xLLR = 2.925
• 4xLLR = 2.935

[paulunderwood]

(2.935*3186)/(4.153*2400) is about 0.938. I am guessing it is because of your superior hardware and higher clocked RAM that you get an extra 6.2% CPU-clock for CPU-clock. If this is error free, using all four cores, you could do a current "321" test in 2 and 3/4 hours per core or, equivalently, in total, 34.6 tests per day However, from experience and from what others say, I do not encourage overclocking.