Should hyperthreading be used?

Electrolyte · 2006-10-29, 19:01

Pretty simple question. If a CPU on a machine supports HyperThreading, should we run 2 clients to take advantage of it?

JuanTutors · 2006-10-29, 19:17

Quote:

Originally Posted by Electrolyte

Pretty simple question. If a CPU on a machine supports HyperThreading, should we run 2 clients to take advantage of it?

Maybe!!!

Actually some people have noticed a couple percent increase in throughput, while others have noticed the same decrease.

If you are just doing LL testing, either just run one copy or try running two copies and see which is faster. Just take the following:
1/(seconds per iteration with just one copy)
1/(seconds per iteration of first copy) + 1/(seconds per iteration of second copy)

Whichever is bigger, take that one.

However, if you're doing two different types of tests, like LL testing and factoring, go ahead and run two copies.

Electrolyte · 2006-10-29, 19:20

I've got the average of LL testing with just one running, so I'll wait to see how it goes when it's running 2 (second client is doing the beginning test at the moment).

kaeptn_kork · 2006-11-03, 03:22

I wouldn't advise running HT with p95, it slows down your normal applications even when running 1 client.
Additionally, running 2 clients on a HT-machine will not speed up crunching at all.

Ethan Hansen · 2006-11-06, 21:19

I wouldn't advise running HT with p95, it slows down your normal applications even when running 1 client.

True enough. When you run a single process of your own that consumes the entire CPU, the Windows process scheduler still sees an unused CPU. It happily lets Prime95 run on the second, hyperthreaded, CPU. The result is that both Prime95 and your process get equal amounts of real CPU time, even though P95 runs at low priority. This brings even the fastest computer to a relative crawl.

There are two approaches to dealing with the situation. You can disable Hyperthreading entirely. If none of the applications you use support multiple process threads, and you only run a single application at a time, this will give the best overall performance. If you do run multiple apps at once or your programs utilize more than one execution thread, George created version 24.15 of Prime95.

V24.15 takes a brute-force approach to the problem. If iteration times decrease slightly, Prime95 pauses for 30 seconds. The algorithm is very sensitive -- so much so that the program pauses for large portions of the day if the computer is being used for anything else. I see anywhere between a 20 and 80% reduction in total P95 throughput depending on how the computer is being used for actual work. There is also a risk of data loss. The code appears to call a sleep routine for 30 seconds, without checking for any Windows messages. Given the choices of disabling hyperthreading, not running Prime95,
having an unusably slow computer, or sacrificing P95 throughput, I chose the latter.

Maybeso · 2006-11-08, 01:29

I've been running two clients on my HT P4 for more than a year. One does LL and the other factoring. The LL throughput drops to around 55%, and the factoring drops to 70%+. Since a P4 is 'better' for LL and there are probably slower machines who would prefer to factor my exponents instead of larger ones, I only run the factoring thread for 10 hours during the day.

On weekends when I use the machine, I've discovered that most of the time I only slow down one of the clients - depending on what app I run. Usually it's the factoring one, which is ok by me. When I run R-Factor (the latest generation of extremely realistic racing sims) or Scorch-3d (realistic 3d artillary) the factoring thread all but stops and the LL thread slows to 20%.

The one problem I do have is when the LL client tries to do TF or P-1 on an exponent before starting the LL. Then both threads slow way down until I notice and can disable the factoring thread for 1 day.

2006-10-29, 19:01	#1
Electrolyte Oct 2006 England 2² Posts	Should hyperthreading be used? Pretty simple question. If a CPU on a machine supports HyperThreading, should we run 2 clients to take advantage of it?

2006-10-29, 19:20	#3
Electrolyte Oct 2006 England 2² Posts	I've got the average of LL testing with just one running, so I'll wait to see how it goes when it's running 2 (second client is doing the beginning test at the moment). Last fiddled with by Electrolyte on 2006-10-29 at 19:20

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Hyperthreading	TheMawn	Hardware	12	2013-08-15 00:03
Hyperthreading	Primeinator	Information & Answers	13	2010-05-20 15:15
Hyperthreading	Jud McCranie	Information & Answers	11	2009-03-05 06:41
Hyperthreading	dave_0273	Hardware	5	2003-12-12 13:22
Hyperthreading	Prodigious	Software	4	2002-12-17 12:31

2006-11-03, 03:22	#4
kaeptn_kork Sep 2002 30₈ Posts	I wouldn't advise running HT with p95, it slows down your normal applications even when running 1 client. Additionally, running 2 clients on a HT-machine will not speed up crunching at all.

2006-11-06, 21:19	#5
Ethan Hansen Oct 2005 101000₂ Posts	I wouldn't advise running HT with p95, it slows down your normal applications even when running 1 client. True enough. When you run a single process of your own that consumes the entire CPU, the Windows process scheduler still sees an unused CPU. It happily lets Prime95 run on the second, hyperthreaded, CPU. The result is that both Prime95 and your process get equal amounts of real CPU time, even though P95 runs at low priority. This brings even the fastest computer to a relative crawl. There are two approaches to dealing with the situation. You can disable Hyperthreading entirely. If none of the applications you use support multiple process threads, and you only run a single application at a time, this will give the best overall performance. If you do run multiple apps at once or your programs utilize more than one execution thread, George created version 24.15 of Prime95. V24.15 takes a brute-force approach to the problem. If iteration times decrease slightly, Prime95 pauses for 30 seconds. The algorithm is very sensitive -- so much so that the program pauses for large portions of the day if the computer is being used for anything else. I see anywhere between a 20 and 80% reduction in total P95 throughput depending on how the computer is being used for actual work. There is also a risk of data loss. The code appears to call a sleep routine for 30 seconds, without checking for any Windows messages. Given the choices of disabling hyperthreading, not running Prime95, having an unusably slow computer, or sacrificing P95 throughput, I chose the latter.

2006-11-08, 01:29	#6
Maybeso Aug 2002 Portland, OR USA 112₁₆ Posts	I've been running two clients on my HT P4 for more than a year. One does LL and the other factoring. The LL throughput drops to around 55%, and the factoring drops to 70%+. Since a P4 is 'better' for LL and there are probably slower machines who would prefer to factor my exponents instead of larger ones, I only run the factoring thread for 10 hours during the day. On weekends when I use the machine, I've discovered that most of the time I only slow down one of the clients - depending on what app I run. Usually it's the factoring one, which is ok by me. When I run R-Factor (the latest generation of extremely realistic racing sims) or Scorch-3d (realistic 3d artillary) the factoring thread all but stops and the LL thread slows to 20%. The one problem I do have is when the LL client tries to do TF or P-1 on an exponent before starting the LL. Then both threads slow way down until I notice and can disable the factoring thread for 1 day.