I know this has been discussed in the Number Crunching thread some, but I think I'm seeing it specifically w/ Cullen WUs. Does Cullen suffer from the cache miss issue with Intel Core i7 with HT enabled? WUs on my i5-2400 system are taking much less time than on my i7-3820 w/ HT.

i5 WU: http://www.primegrid.com/workunit.php?wuid=336493532 (112k sec)
i7 WU: http://www.primegrid.com/workunit.php?wuid=336464178 (158k sec)
____________
My lucky #: 60133106^131072+1 (GFN 17-mega)

HT is just enabled or HT is fully used?
____________
92*10^^1585996-1 NEAR-REPDIGIT PRIME :) :) :)
4 * 650^⁴⁹⁸¹⁰¹-1 CRUS PRIME
2022202116^¹³¹⁰⁷²+1 GENERALIZED FERMAT
Proud member of team Aggie The Pew. Go Aggie!

Fully used. I'm running on all 8 cores.

An aside, how would I configure PG to not use *all* of the HT cores, but partial? Set to 3/4 CPUs (and thus 6/8 cores?)
____________
My lucky #: 60133106^131072+1 (GFN 17-mega)

I know this has been discussed in the Number Crunching thread some, but I think I'm seeing it specifically w/ Cullen WUs. Does Cullen suffer from the cache miss issue with Intel Core i7 with HT enabled? WUs on my i5-2400 system are taking much less time than on my i7-3820 w/ HT.

i5 WU: http://www.primegrid.com/workunit.php?wuid=336493532 (112k sec)
i7 WU: http://www.primegrid.com/workunit.php?wuid=336464178 (158k sec)

i5 WU: http://www.primegrid.com/workunit.php?wuid=336493532 (112k sec)
i7 WU: http://www.primegrid.com/workunit.php?wuid=336464178 (158k sec)


This is great result. You belong to only few peoples that have increased performance with HT = ON . Many others got drop in performance...
In your case

112000 /4 = 28000 per WU
158000 /8 = 19750 per WU.
So ideally will be 14000 per WU, but you result is near perfect. As long as you with HT enabled compute below 56000 per WU you have performance increased...
____________
92*10^^1585996-1 NEAR-REPDIGIT PRIME :) :) :)
4 * 650^⁴⁹⁸¹⁰¹-1 CRUS PRIME
2022202116^¹³¹⁰⁷²+1 GENERALIZED FERMAT
Proud member of team Aggie The Pew. Go Aggie!

you should probably find it's faster per wu to use only 6/8.

You belong to only few peoples that have increased performance with HT = ON .

you should probably find it's faster per wu to use only 6/8.

You belong to only few peoples that have increased performance with HT = ON .

have you tried for yourself to see what option gives the best times per wu?

you should probably find it's faster per wu to use only 6/8.

You belong to only few peoples that have increased performance with HT = ON .

have you tried for yourself to see what option gives the best times per wu?

I may try later, to see if 3/4 CPUs (6/8 cores) gives me better time, but it's not a big deal. The only thing I really note is that when running 8/8 along w/ a PPS Sieve task, one of the Cullen (or any CPU-based) WUs suffers accordingly.

darkclown's i5 and i7 are not comparable, simply by measuring HT "on" vs. "off". The i5-2400 is "mid range", while the i7-3820 is "extreme", with 0.5G faster clocks and 4GB more L3 cache.

I ran another test. This is an extension of a test I made earlier this month comparing the cache-miss effect on Sandy vs. Ivy. I posted about this on the AtP message board. At that time I was comparing the two processors, and not worrying about hyperthreading ("off" for all tests). Bottom line was that Ivy seemed more severely impacted under heavy load.

This time, I'm comparing HT "off" vs. HT "on", on the Ivy Bridge 3770K. I used prpnet instances to run the exact same Woodall number (11574717*2^11574717-1, which I had earlier run in boinc and had validated) 1 through 8 times in parallel (just to a few percent, long enough to let the timing stabilize). Nothing else going on on the box, not even GPU. These are the "time per iteration" values reported by LLR (3.8.9), so lower is faster (units ms.). Ubuntu 12.04.

HT off:
1: 6.239
2: 6.657
3: 7.789
4: 9.940

HT on:
1: 6.229
2: 6.634
3: 7.649
4: 9.712
5: ~12.9
6: ~15.0
7: ~18.2
8: 20.624

The times for 5-7 are rougher averages over the various LLR instances, as you never know at any given point which one will be forced to "share", and which will have a dedicated core. The scheduler keeps them pretty well balanced when there are 4 or fewer, or 8.

Of course remember this is one machine, running one number. Your mileage may vary.

--Gary

If I read that correctly it shows that 6 cores is best for you too.

If I read that correctly it shows that 6 cores is best for you too.

In sieving my I7-2700K got about 18% increase when using HT
____________
92*10^^1585996-1 NEAR-REPDIGIT PRIME :) :) :)
4 * 650^⁴⁹⁸¹⁰¹-1 CRUS PRIME
2022202116^¹³¹⁰⁷²+1 GENERALIZED FERMAT
Proud member of team Aggie The Pew. Go Aggie!

If I read that correctly it shows that 6 cores is best for you too.

9.94/4=2,485 per WU when HT is off
20.624/8= 2.578 per WU when HT is on.

So what I see that result with HT is always slower. Even sweet spot of 6 WU (with HT) have average 2.5 vs 2,485 (without HT)

In that case, you are right :)
Since I do just LLR-ing I put HT off :)
____________
92*10^^1585996-1 NEAR-REPDIGIT PRIME :) :) :)
4 * 650^⁴⁹⁸¹⁰¹-1 CRUS PRIME
2022202116^¹³¹⁰⁷²+1 GENERALIZED FERMAT
Proud member of team Aggie The Pew. Go Aggie!