2CPU.com » News » January 2004 » Hyper-threading Performance Explored: Part I

As I mentioned earlier today, I've been working on another article detailing hyper-threading performance. This time around, I wanted to get a little more technical and possibly even look to the future as well. I decided to break up the article into two separate parts: In part I, we'll be dealing with Windows performance, and later in Part II, I'll be digging into some Linux 2.6 benchmarks.

... it's important (and somewhat obvious) to note that a CPU that supports hyper-threading is not going to provide comparable performance with two physical processors rated at the same speed. The simple reason for this is because the two logical processors that make up your hyper-threaded CPU have to share resources, namely the execution engine, cache, and access to the system bus.

If you're interested in hyper-threading at all, you'll probably want to take a look at this.

05/06/2003 02:36 AM: The Proliferation of Hyper-Threading: Intel's Pentium 4 'C' by Jim
Awhile ago we took a look at Hyperthreading performance using 2.0GHz and 2.8GHz Xeons. Since that time Intel has of course brought HT to the desktop processor line as well. Has the technology improved...

11/14/2002 08:44 AM: Pentium 4 3.06 GHz with Hyper-Threading Reviewed at TR by Cory
It looks like the NDAs on Intel's new 3.06 GHz Pentium 4 expire today, and our hommies at TR are the first ones out the gate with their review of the first desktop P4 to feature Hyper-Threading. Quot...

ReMeDy Posts: 12521 Joined: 2001-02-10

#25690 Posted on: 01/07/2004 10:38 AM Nice work. |[size=1][url=http://forums.2cpu.com/showthread.php?s=&postid=279831#post279831][color=yellow]FAQ's[/url][/color][/size]|Heatware-Remedy{WcS}|Dual760|Beerology|

nipster NOT a WarMonger Posts: 1841 Joined: 2001-09-12

#25691 Posted on: 01/07/2004 10:51 AM something you may want to try in the future is how HT scales with DC projects like seti and folding, etc... i would be interested to know im kind of running HT blindly with 2 seti processes on my single P4C 2.8, never ran it with a single process since upgrading from the 2.66 non-ht cpu dont know how much of a hit or gain i am getting

EM NOT a WarMonger Posts: 4111 Joined: 2000-09-04

#25692 Posted on: 01/07/2004 11:16 AM Very interesting read....good work. Political Correctness is a doctrine, fostered by a delusional, illogical, liberal minority and rabidly promoted by an unscrupulous mainstream media, which holds forth the proposition that it is entirely possible to pick up a turd by the clean end. Artificial intelligence is no match for natural stupidity. Heat

AssKoala Anti-Zealot @ GATech Posts: 3309 Joined: 2002-01-02

#25693 Posted on: 01/07/2004 11:36 AM Nice, good job. Me Webpage | If you always think like an expert, you'll always be a beginner. | "A handful of knowledgeable people is more effective than an army of fools" -Writing Secure Code, 2nd Ed.

jhaislet "Let's Roll" Posts: 3257 Joined: 2002-06-25

#25694 Posted on: 01/07/2004 11:42 AM Great work Jim! We look forward to the second part and hope your Intel source can get you the "hook-up" on your requested parts. Mini-ITX Systems: AOpen MP965-DR Core2Duo Mobility T7500 4MB L2, 4GB DDR2 (Vista MCE)| MSI Industrial 945GM2 Celeron M 430 | Intel Little Valley 2 (Firewall) | Via EPIA EN-1500 (Firewall)|

ayt999 "Let's Roll" Posts: 3038 Joined: 2002-10-28

#25695 Posted on: 01/07/2004 12:03 PM that was a nice read, now I'm waiting for part II. keep up the good work. SETI@home|folding@home|RC5|find-a-drug|DPAD|ZetaGrid|climateprediction.net| communityTSC|BOINC|Predictor@home|seventeen or bust|ECC2-109

spikegifted I do what? Posts: 6144 Joined: 2000-07-24

#25696 Posted on: 01/07/2004 02:42 PM Nice work, Jim_!! Originally posted by nipster something you may want to try in the future is how HT scales with DC projects like seti and folding, etc... i would be interested to know Just echoing nipster, I'm kinda interested to see what kind of performance boost/loss between say 2 S@H clients with dual Xeon without HT and 4 S@H clients with dual Xeon with HT enabled... spikegifted.net "At the very beginning of the 19th century chemists generally thought that compounds from living organisms were too complicated in structure to be capable of artificial synthesis from non-living things, and that a 'vital force' or vitalism conferred the characteristics of living beings on this form of matter." Well, how wrong were they? So much for dogma...

815SMP i815 SMP Retro-Rig Posts: 743 Joined: 2003-09-30

#25697 Posted on: 01/07/2004 02:57 PM Great read, now we know what you did over the Holidays!

NerdZero Registered User Posts: 199 Joined: 2003-03-11

#25698 Posted on: 01/07/2004 03:55 PM Originally posted by spikegifted Nice work, Jim_!! Just echoing nipster, I'm kinda interested to see what kind of performance boost/loss between say 2 S@H clients with dual Xeon without HT and 4 S@H clients with dual Xeon with HT enabled... I have to third nipster's suggestion. I'm really interested in seeing on how thread priorities affect hyperthreading performance. I know low priority apps will cause a slow down on higher priority ones, just wondering how much of a slowdown. Should us SETI crunchers have hyperthreading disabled to get the most out of our system? Any thoughts? Great work _Jim! Thanks for taking the time to get some numbers for us.

Jim_ Administrator Posts: 3464 Joined: 2000-03-15

#25699 Posted on: 01/07/2004 06:37 PM I do believe it was one of our slashdot readers that was kind enough to let me know that I miscalculated a couple of the percentages. I fixed the couple of errors and updated the article. Thanks Fred! [url="http://www.2cpu.com"][size=1]2CPU.com[/url] - Because two are always better than one! [url="http://www.jimkirk.org"]jimkirk.org[/url] - Not a Myth any Longer. Just a Dad.[/size]

big boi fullblown SMP addict Posts: 1382 Joined: 2002-02-13

#25700 Posted on: 01/07/2004 07:49 PM can someone please explain the math behind this statement to me: "Enabling HT on the Xeons gets us to about a whopping 110% improvement in performance." it went from 519 to 591 unless i'm missing what you were comparing jim. or i just can't understand the math- entirely possible. other than that i thought it was great. very informative article jim. i would also like to say that i'm just giddy at the way things seem to be going for you guys in the new year. news posts and articles seemed to be pretty dead at the end of 2003. this week you guys have put up a lot of content, let us know about planned articles and reviews, your contacts with the hardware industry seem to be really coming together, and as we all saw with this article- when you guys get your hands on the good stuff you always come through with excellent content. looking forward to part 2 and everything else 2004 has to bring us lucky 2cpu followers big boi A great source for computer and communications news: www.geekextreme.com heatware ebay

Jim_ Administrator Posts: 3464 Joined: 2000-03-15

#25701 Posted on: 01/07/2004 07:50 PM Originally posted by big boi can someone please explain the math behind this statement to me: "Enabling HT on the Xeons gets us to about a whopping 110% improvement in performance." My fault, I was juggling some numbers this morning and I just WORDED that improperly. I'll fix it. It should say something to the effect of: "When scaling from one processor to multiple processors WITH hyper-threading enabled, we see a performance improvement of approximately 110%." [url="http://www.2cpu.com"][size=1]2CPU.com[/url] - Because two are always better than one! [url="http://www.jimkirk.org"]jimkirk.org[/url] - Not a Myth any Longer. Just a Dad.[/size]

big boi fullblown SMP addict Posts: 1382 Joined: 2002-02-13

#25702 Posted on: 01/07/2004 07:54 PM yeah, i didn't think it was actually a math issue- just that i didn't know what the 2 points of comparison were. A great source for computer and communications news: www.geekextreme.com heatware ebay

Forge fullblown SMP addict Posts: 704 Joined: 2001-05-12

#25703 Posted on: 01/07/2004 11:59 PM WRT: HT and S@H: It's not good. I don't have numbers handy, but on my 2.8C, S@H performance, from best to worst, was: one client, HT off one client, HT on two clients, HT on two clients, HT off S@H is just harsh on L2/FSB/memory bandwidth. I've heard the working set averages around 1MB, so if Prescott EE has 2MB L2 or more, then it might make sense to run two clients+HT. Once I get some RAM for my 2.8C rig, I'll run some dnetc/S@H numbers for ya. Folding@Home, unfortunately, just isn't easy to bench, with the work units being so variable and even the crunching engines changing from WU to WU. Edit: The GUI S@H client seemed to run much more smoothly with HT on and a single copy, now that I think back. I doubt any serious SETI freaks are using the GUI client, though. Registered Linux user 82133 (li.org has a short memory)

jhaislet "Let's Roll" Posts: 3257 Joined: 2002-06-25

#25704 Posted on: 01/08/2004 12:27 AM Congrat's Jim, The Inquire picked up your article and posted a link. http://www.theinquirer.net/?article=13486 Mini-ITX Systems: AOpen MP965-DR Core2Duo Mobility T7500 4MB L2, 4GB DDR2 (Vista MCE)| MSI Industrial 945GM2 Celeron M 430 | Intel Little Valley 2 (Firewall) | Via EPIA EN-1500 (Firewall)|

Augustus Registered User Posts: 315 Joined: 2002-10-29

#25705 Posted on: 01/08/2004 01:04 AM Jim, On the last page of the article you state that the Pentium 4 EE has 1MB of L3 cache. Actually, it has 2MB of L3. Check out the Intel press release here. Good article. Lots of good info. Looking forward to the Linux half. Augustus

Jim_ Administrator Posts: 3464 Joined: 2000-03-15

#25706 Posted on: 01/08/2004 01:22 AM Originally posted by Augustus Jim, On the last page of the article you state that the Pentium 4 EE has 1MB of L3 cache. Actually, it has 2MB of L3. Check out the Intel press release here. Good article. Lots of good info. Looking forward to the Linux half. Whoops! Thanks Augustus. [url="http://www.2cpu.com"][size=1]2CPU.com[/url] - Because two are always better than one! [url="http://www.jimkirk.org"]jimkirk.org[/url] - Not a Myth any Longer. Just a Dad.[/size]

Jim_ Administrator Posts: 3464 Joined: 2000-03-15

#25707 Posted on: 01/08/2004 04:01 AM You know, I think the best thing about us getting slashdotted is always that some of the people who breeze through the article swing by the forums to check them out. If they spend any amount of time here, you just know they're going to want to come back. As I post this there are 90+ people on the forums, and I think it's great. [url="http://www.2cpu.com"][size=1]2CPU.com[/url] - Because two are always better than one! [url="http://www.jimkirk.org"]jimkirk.org[/url] - Not a Myth any Longer. Just a Dad.[/size]

ReMeDy Administrator Posts: 12521 Joined: 2001-02-10

#25708 Posted on: 01/08/2004 05:00 AM Originally posted by Jim_ You know, I think the best thing about us getting slashdotted is always that some of the people who breeze through the article swing by the forums to check them out. If they spend any amount of time here, you just know they're going to want to come back. As I post this there are 90+ people on the forums, and I think it's great. There are currently 36 member(s) and 50 guest(s) on the boards. Most users ever online was 172 on 05-11-2001 at 12:16 PM :eek: :eek: :eek: |[size=1][url=http://forums.2cpu.com/showthread.php?s=&postid=279831#post279831][color=yellow]FAQ's[/url][/color][/size]|Heatware-Remedy{WcS}|Dual760|Beerology|

mekboy Registered User Posts: 748 Joined: 2001-11-18

#25709 Posted on: 01/08/2004 05:04 AM You were linked on HardOCP. That alone will get you hosed.

widefault Registered User Posts: 1202 Joined: 2001-02-06

#25710 Posted on: 01/08/2004 05:37 AM Originally posted by Forge WRT: HT and S@H: It's not good. I don't have numbers handy, but on my 2.8C, S@H performance, from best to worst, was: one client, HT off one client, HT on two clients, HT on two clients, HT off HT does slow each individual client, but you're not taking into account the increase in production per physical CPU HT offers. I'm getting around 30% more WUs a day using HT and two clients compared to just one client.

spikegifted I do what? Posts: 6144 Joined: 2000-07-24

#25711 Posted on: 01/08/2004 05:51 AM Originally posted by widefault I'm getting around 30% more WUs a day using HT and two clients compared to just one client. Is that with a P4 or a pair of Xeons? With my Xeons I'm running 4 clients concurrently... spikegifted.net "At the very beginning of the 19th century chemists generally thought that compounds from living organisms were too complicated in structure to be capable of artificial synthesis from non-living things, and that a 'vital force' or vitalism conferred the characteristics of living beings on this form of matter." Well, how wrong were they? So much for dogma...

spikegifted I do what? Posts: 6144 Joined: 2000-07-24

#25712 Posted on: 01/08/2004 05:52 AM Originally posted by mekboy You were linked on HardOCP. That alone will get you hosed. Interestingly, the SMP section of the [H] forum is probably the most civilized part... (All thanks to a good number of 2CPUers spending time there to teach them a thing or two! ) spikegifted.net "At the very beginning of the 19th century chemists generally thought that compounds from living organisms were too complicated in structure to be capable of artificial synthesis from non-living things, and that a 'vital force' or vitalism conferred the characteristics of living beings on this form of matter." Well, how wrong were they? So much for dogma...

widefault Registered User Posts: 1202 Joined: 2001-02-06

#25713 Posted on: 01/08/2004 07:09 AM Originally posted by spikegifted Is that with a P4 or a pair of Xeons? With my Xeons I'm running 4 clients concurrently... P4 2.4C@3GHz on an i865 board.

Ytterbium SMP Newbie Posts: 1 Joined: 2004-01-08

#25714 Posted on: 01/09/2004 01:30 AM Nice article, caused me to register and post this! the EE version of the P4 doesn't really do much in DVD encoding the extra 1.5Mb of cache may not be used by TMPEnc?? see here http://akiba.ascii24.com/akiba/news/2003/10/25/imageview/images727231.gif.html http://www.computerbase.de/artikel/hardware/prozessoren/intel_pentium_4_extreme_edition_2_mb_cache/11/ http://www.hardwareluxx.de/cgi-bin/ubb/ultimatebb.cgi?ubb=get_topic&f=3&t=000010&p=9 to me it's always seemed to be Mhz that count's as the AMD CPU's always do poorly in TMPEnc? I suppose my read of Mhz maybe FSB although this is a bit strange as 3.06 beat's 3.0 http://www.beareyes.com/2/lib/200304/15/20030415240.htm http://www.beareyes.com/2/lib/200304/15/240/4.gif I'm sure there's alot of inconsistancy in the way the tmpenc benches have been done though!!

2CPU.com » News » January 2004 » Hyper-threading Performance Explored: Part I