Message boards :
Number crunching :
Pentium-M Linux Performance?
Message board moderation
Author | Message |
---|---|
TonyM Send message Joined: 13 Dec 99 Posts: 6 Credit: 485,503 RAC: 0 |
I have noticed that in Linux, my 1.6GHz Pentium-M (Dothan) will take about 8Hrs to complete a unit, my 400MHz Celeron system, which only uses linux, takes 17Hrs. So the Pentium-M is only twice as fast?! However, when in Windows it takes about 2.5Hrs to complete a work unit. What the heck is going on here? Why is there such a larger margin? |
Zeeno Send message Joined: 19 May 00 Posts: 20 Credit: 75,268 RAC: 0 |
I don't use that cpu but there are a few guys on my team that do. One of them may be able to compare findings with you or answer your question. http://www.team-scifi.com. Good luck. > I have noticed that in Linux, my 1.6GHz Pentium-M (Dothan) will take about > 8Hrs to complete a unit, my 400MHz Celeron system, which only uses linux, > takes 17Hrs. So the Pentium-M is only twice as fast?! > > However, when in Windows it takes about 2.5Hrs to complete a work unit. > > What the heck is going on here? Why is there such a larger margin? > |
mikey Send message Joined: 17 Dec 99 Posts: 4215 Credit: 3,474,603 RAC: 0 |
I have several computers crunching for Boinc and 1 is a 1.6 desktop. It takes about 5 hours to crunch a unit. I use Windows and have 2000 on it. My AMD 64 3200+ is taking about 2 hours to complete a Boinc unit. I also have an old server crunching Classic units, it is a dual 300mhz system that does a unit in almost 19 hours. It is doing 2 at the same time and is running Windows 2000 Server. I think your 400 doing a unit in 2.5 hours is not normal, the unit must have finished early due to any of a number of reasons. The Pentium-M computer, I guess it is a laptop, may not have a full 512k of L2 cache. It could be a Celeron chip meaning that the 384k client will not fit totally into the L2 cache making the unit swap constantly therby slowing it down ALOT! I also have 2 Celeron system crunching for Boinc, the first is a 2.5ghz that does units in almost 8 hours, the second is a 2.4ghz system that is doing units in almost 8.5 hours. Another system I have crunching is an AMD 2200+ and it is doing Boinc units in about 3 hours 15 minutes. Click on my name on the left and then on computers/view and you will see all of my computers crunching for Boinc. Just the one with the IP address is not currently crunching for Boinc, it is back to crunching Classic units. |
Hans Dorn Send message Joined: 3 Apr 99 Posts: 2262 Credit: 26,448,570 RAC: 0 |
> I have noticed that in Linux, my 1.6GHz Pentium-M (Dothan) will take about > 8Hrs to complete a unit, my 400MHz Celeron system, which only uses linux, > takes 17Hrs. So the Pentium-M is only twice as fast?! > > However, when in Windows it takes about 2.5Hrs to complete a work unit. > > What the heck is going on here? Why is there such a larger margin? > Hmmm. Sounds strange. I'm getting a little below 3hrs with debian and a stock seti client. Could you post the contents of your "/proc/cpuinfo"? I'd like to check if your CPU is running at the proper speed. Regards Hans |
Ned Slider Send message Joined: 12 Oct 01 Posts: 668 Credit: 4,375,315 RAC: 0 |
That does sound slow. As it's a Pentium-M aimed at the mobile market, I wonder if it's running some kind of throttling (at reduced speed). Ned *** My Guide to Compiling Optimised BOINC and SETI Clients *** *** Download Optimised BOINC and SETI Clients for Linux Here *** |
TonyM Send message Joined: 13 Dec 99 Posts: 6 Credit: 485,503 RAC: 0 |
Hmm, seems as though I figured out what is causing this :( It seems as though the cpu frequency scaler I use doesn't take into account whether the laptop is plugged in or not :( It was running at 600MHz lol... So, does anyone know of a good cpufreq scaler that takes into account whether the laptop is plugged in or not? Also note, the 2.5Hrs is for the Pentium-M in windows, not the 400MHz Celeron ;) |
Paul D. Buck Send message Joined: 19 Jul 00 Posts: 3898 Credit: 1,158,042 RAC: 0 |
> Also note, the 2.5Hrs is for the Pentium-M in windows, not the 400MHz Celeron The Celeron is also much more crippled than some of its specs might indicate. I can't keep up with all of the architectures like I used to, but the Celeron if memory serves has both a smaller cache and also has changes to the pipeline both of which will impac performance. |
Hans Dorn Send message Joined: 3 Apr 99 Posts: 2262 Credit: 26,448,570 RAC: 0 |
> Hmm, seems as though I figured out what is causing this :( > > It seems as though the cpu frequency scaler I use doesn't take into account > whether the laptop is plugged in or not :( It was running at 600MHz lol... > > So, does anyone know of a good cpufreq scaler that takes into account whether > the laptop is plugged in or not? > Do you have cpufrequtils? It contains a tool called "cpufreq-set" that should allow you setting the frequency manually. I can't try this here, because I don't have a laptop myself. For a current version look here: http://www.kernel.org/pub/linux/utils/kernel/cpufreq/cpufrequtils.html Regards Hans |
Roberto Virga Send message Joined: 29 May 99 Posts: 32 Credit: 41,362 RAC: 0 |
If you're running Fedora Core 2-3, or any similar distro which uses cpuspeed, there's a simple explanation for what you're seeing. cpuspeed reads "/proc/stat" periodically to monitor the computer's activity, and scales up or down the cpu frequency according to the system load. The "/proc/stat" file contains 4 key statistics: system time (time spent in system mode), user time (time spent in user mode), nice time (time spent in user mode at a lower priority), idle time. Unfortunately, what happens is that (1) all computation by SETI@home is counted in nice time, and (2) cpuspeed treats nice time the same as idle time. This means that if all you have running is SETI@home, cpuspeed will act like you don't have anything running at all, and it will scale the cpu frequency all the way down to 600MHz. Do "cat /proc/cpuinfo" to verify this. I have a similar configuration (Pentium-M Banias, 1.6GHz), and I had to patch the cpuspeed source to be able to run SETI@home at full speed (while still enjoing the advantages of dynamic CPU frequency scaling in all other circumstances). - Roberto |
TonyM Send message Joined: 13 Dec 99 Posts: 6 Credit: 485,503 RAC: 0 |
>The Celeron is also much more crippled than some of its specs might indicate. I >can't keep up with all of the architectures like I used to, but the Celeron if >memory serves has both a smaller cache and also has changes to the pipeline both >of which will impac performance. I'm not sure what you are refering to, but I was just clarifying for someone else who was confused earlier. Ok, here is my attemp to make it crystal clear. Just for fun, I'll list all my computers, this is averages discounting all units that took significantly less time. CPU - - - - - Frequency - WIN - LINUX (hrs) ------------------------------------------- Pentium-M - - 1.6GHz - - 2.5 - 8* Celeron - - - 400MHz - - na - 17 Pentium 3 - - 1.4GHz - - 7.0 - 7 AthlonXP 2800 2.1GHz - 3.1 - na Athlon64 3200 2.1GHz - 2.2 - 2.2 VIAC3-2 - - - 1.0GHz - - na - 25.5 *seems to have been running at 600MHz |
Paul D. Buck Send message Joined: 19 Jul 00 Posts: 3898 Credit: 1,158,042 RAC: 0 |
> I'm not sure what you are refering to, but I was just clarifying for someone > else who was confused earlier. Ok, here is my attemp to make it crystal clear. > Just for fun, I'll list all my computers, this is averages discounting all > units that took significantly less time. What I was trying to say was that, based solely on clock speed alone, the Celeron will appear slower than it should. In other words, it runs at 400 MHz, theory says it will be 1/2 as fast as a 800 MHz P4 (or AMD equiv.), yet it will be worse than that. Something like 1/4 ... For a long time now, many of us geeks have been trying to explain to the universe that clock speed does not matter. The other contrast is where AMD used to rate their CPUs using an equivelent. So, one of their 1.2 MHz chips might be rated as an Intel 1.4 GHz using the number 1400 ... I used to have a mixed company and the basic AMD numbers compared out based on my experience. So, the long, oft told refrain was true, AMDs were cheaper and could achieve the same results with a slower clock speed. I was having trouble falling asleep last night so I did a little looking and I could get a Xeon with 1M cache for about $700, dual MB would be ~$200 so I could probably upgrade one of my chassis (or buy another one) and build a box that would give me 4 CPUs. Most important would be the fact that I could spread the total purchase over a couple months ... and for a pure crunch box that would not be half bad ... Of course, before I can really do that I would need to get a wireless LAN up with sharing of the Internet connection so that I could move at least 2 of the boxes out of the computer room. Oh, and update the current boxes to Linux. Heck, that wold be a plan! |
Toby Send message Joined: 26 Oct 00 Posts: 1005 Credit: 6,366,949 RAC: 0 |
> The Pentium-M computer, I guess it is a laptop, may not have a full 512k of L2 > cache. It could be a Celeron chip meaning that the 384k client will not fit > totally into the L2 cache making the unit swap constantly therby slowing it > down ALOT! Heh... you couldn't be more wrong on that :) The Pentium-M chips have 1,024 KB of L2 cache. I think newer ones even have 2,048. They are also based on the Pentium 3 core which was MUCH better than any P4 will ever be by the looks of it. My 1.3 GHz Pentium-M eats my 2600+ AMD for breakfast when it comes to seti. The AMD takes about 3.5 hours and the Pentium-M usually takes under 3. Pretty impressive really. And all this while taking up a maximum of about 30 watts of power. (under 10 when idle and throttled down to 600 MHz) A member of The Knights Who Say NI! For rankings, history graphs and more, check out: My BOINC stats site |
shady Send message Joined: 2 Feb 03 Posts: 40 Credit: 2,640,527 RAC: 0 |
The Dothan CPU's are very seti friendly and yes the current versions do have 2mb of L2 cache. I have one of the 1.6 chips running at 2.6 that did our team seti classic benchmark wu in 1hr 12mins. (its yet to do any seti boinc as its on predictor duty at present). They also run nice and cool and quiet , dont need a large power supply and as the boards are m-atx they dont take up much room. When the next generation of desktop motherboards are out for them (will support dual channel ddr2) hopefully a few more manufacturers will join the party and the price of the motherboards will come down a bit. These Dothans currently look like the ideal set up for seti farms :) Shady <img src='http://www.boincsynergy.com/images/stats/comb-1527.jpg'> |
Paul D. Buck Send message Joined: 19 Jul 00 Posts: 3898 Credit: 1,158,042 RAC: 0 |
Shady, > These Dothans currently look like the ideal set up for seti farms :) My problem is that I can never seem to find the "super-duper" what ever named deally bob (sorry for all the technical terms). Oh, well, FIRST we save enough for the next Macintosh top end tower. After that I can think of making another tower for pure BOINC work. |
Toby Send message Joined: 26 Oct 00 Posts: 1005 Credit: 6,366,949 RAC: 0 |
The one thing about the pentium-M chips that I have found is that they give you VERY different milage on different projects. As I said in my last post, my 1.3 GHz Pentium-M beats my AMD 2600+ while doing seti. However, it appears to be lagging FAR behind it on its first Einstein work unit. The 2600+ takes about 7.5 hours for an Einstein work unit but the Pentium-M looks like it will take over 10 hours. It is 27% done after 3:11. I'm guessing that seti@home somehow makes better use of the large L2 cache in the way it uses data from memory whereas some of the other projects work in such a way as to nullify this advantage it has over other CPUs. Maybe if I get REALLY bored I will look at the assembly for the science apps and figure out what is causing this... but I probably won't be THAT bored until early in the next century... A member of The Knights Who Say NI! For rankings, history graphs and more, check out: My BOINC stats site |
Paul D. Buck Send message Joined: 19 Jul 00 Posts: 3898 Credit: 1,158,042 RAC: 0 |
> The one thing about the pentium-M chips that I have found is that they give > you VERY different milage on different projects. As I said in my last post, > my 1.3 GHz Pentium-M beats my AMD 2600+ while doing seti. However, it appears > to be lagging FAR behind it on its first Einstein work unit. The 2600+ takes > about 7.5 hours for an Einstein work unit but the Pentium-M looks like it will > take over 10 hours. It is 27% done after 3:11. I'm guessing that seti@home > somehow makes better use of the large L2 cache in the way it uses data from > memory whereas some of the other projects work in such a way as to nullify > this advantage it has over other CPUs. Maybe if I get REALLY bored I will > look at the assembly for the science apps and figure out what is causing > this... but I probably won't be THAT bored until early in the next century... I suspect that it is that the Einstein@Home WU is larger and therefore causes thrashing in the cache, where the SETI@Home WU can fit into the cache so you always get hits. When I look at my EAH dirs I see over 11M for the data files and application. |
Ned Slider Send message Joined: 12 Oct 01 Posts: 668 Credit: 4,375,315 RAC: 0 |
> If you're running Fedora Core 2-3, or any similar distro which uses cpuspeed, > there's a simple explanation for what you're seeing. > cpuspeed reads "/proc/stat" periodically to monitor the computer's activity, > and scales up or down the cpu frequency according to the system load. The > "/proc/stat" file contains 4 key statistics: system time (time spent in system > mode), user time (time spent in user mode), nice time (time spent in user mode > at a lower priority), idle time. > Unfortunately, what happens is that (1) all computation by SETI@home is > counted in nice time, and (2) cpuspeed treats nice time the same as idle time. > This means that if all you have running is SETI@home, cpuspeed will act like > you don't have anything running at all, and it will scale the cpu frequency > all the way down to 600MHz. Do "cat /proc/cpuinfo" to verify this. > I have a similar configuration (Pentium-M Banias, 1.6GHz), and I had to patch > the cpuspeed source to be able to run SETI@home at full speed (while still > enjoing the advantages of dynamic CPU frequency scaling in all other > circumstances). > > - Roberto > That would make a lot of sense. I'm still running Fedora Core 1 on my laptop and it runs boinc/seti at full speed, although it's a Mobile Athlon XP 1400+ that runs at either 1200MHz (12x100) or 500MHz (5x100) in power saving mode. Ned *** My Guide to Compiling Optimised BOINC and SETI Clients *** *** Download Optimised BOINC and SETI Clients for Linux Here *** |
©2024 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.