Yep, I know it's not Windows as such which is to blame. I dual booted before, with different projects or app versions, and never came across more than maybe 5% difference, one way or the other. I think some CPUs just react adversely to the current science app using the Windows code. Strange that your Pent M doesn't, though. Seeing the performance increase on both the P3 and the Core I thought it would. Yeah well, worst case one has to try it out for one's own CPU if there are no reliable rules about it. Still, I would appreciate some ideas from the project staff about the possible reasons. Bernd, are you reading this? Could you perhaps share some ideas with us curious volunteers? ;-)
Quote:
So, will you continue to run that host under Linux? 30% is quite a difference.
Actually, I think I will. The new Ubuntu has some really nice programs on board, so I have no problems working under Linux, and most of the important hardware (graphics, ACPI, dual core support, WLAN, even the function keys*) works fine either out of the box or with some tweaking. This product series of Samsung notebooks seems to be suited well for Linux, especially Debian/Ubuntu, from what you read on the net. I'm also confident to get my IPSec up and running quite easily, or at least I'll know who to ask, so, no great problems there (I hope). My sound is annoying me a bit atm (meaning I don't have any), which seems to be a rather common bug with Feisty Fawn, but I'll force that to cooperate sooner or later.
So, it looks like there are not many arguments against running Linux as "primary" OS on this box. As you said, 30% more performance are a huge difference. It would mean I would finish a day's share of WUs in about 17 hours ;-) not bad, isn't it?
*Hope that is the correct English term. I mean when you press "Fn" and an F key and have features like switch off the sound or display or activate WLAN.
Sounds like finally running Linux on a notebook is a real alternative now. I was never very convinced about WLAN support and things like power saving (dynamic CPU frequency switching).
Anyway...I'll try to crunch some more units under XP and wait and see what happens. We've seen strange things happening when you allow the graphics to display(slowdown even after closing the graphics again), it must be some equally strange effect that eats up CPU cycles on some systems, and not on others.
It would be cool if others who suspect a poor performance under this or that OS could try to run a few units on the same host but with a different OS (Live-CD DVD Linux distributions come in handy here).
Enjoy the rest of the holiday, too bad you have to study tomorrow (sorry for reminding you :-) :-) :-) :-) ).
Actually, I'm studying now ;-) as the code for tomorrow won't prepare itself and I had absolutely no time to do it earlier. Gotta get used to it I guess, the IT business is not exactly "nine to five" either...
Yeah, Linux on a laptop works okay I think. CT suspects about 5% less battery life, because power saving is not quite as efficient yet, but I think I can live with that atm. Still not figured out my sound problems, though...
Surprisingly, on my particular box, there's no difference whatsoever so far between Linux and Win XP performance in terms of credits/CPU hr.
Sorry for not updating my progress but this difference between Linux and Windows on Intel hardware is turning into quite a big deal for me. I'm up to five separate installations now (and a sixth on the way) and all of them are showing a significant improvement for the 4.21 Linux app.
Quote:
It's a Pentium M which (in the development history of Intel CPUs), is kind of similar to both it's architectural predecessor (Pentium III, where Gary found a significant performance delta) and it's successor (Core, as in your host).
So far I've done 2 x Coppermine dual PIII 866 Xeons, 1 x Coppermine dual PIII 866, 1 x Tualatin dual PIII 1266 (the S version with double L2 cache), and 1 x Tualatin single PIII 1200 (standard L2 cache). The sixth install is a P4 Williamette 1.7GHz as I want to see how the early (and terrible!!) P4 architecture behaves.
Only one machine has actually had results returned and validated so far but all of the others are between 25% and 75% complete and this is enough to start making predictions. Assuming that the team have done the best job possible in matching allocated credit to the actual work performed, I think that credit per cpu hour (CCH) is a worthwhile yardstick to use for comparing crunching efficiency of the two operating systems. My own observation is that CCH is remarkably uniform for a given machine/OS combination and relatively independent of the actual result being crunched.
There is quite an architecture change between Coppermine PIIIs and Tualatin PIIIs which is why I was keen to test out some Tualatins. The Tualatin PIII had a quite short lifecycle as Intel killed it off to make way for the Williamette (and then Northwood) P4s. Once again Marketing won at the expense of good engineering :). The Tualatin was later resurrected and further improved to become the Pentium M.
As mentioned previously, I intend to write my findings into a separate thread but with the expansion of installs, that will certainly be a little later on :). In the meantime here is a summary of preliminary results - note C means Coppermine and T means Tualatin and the % increase is for the 4.21 Linux app as compared to the 4.17 windows app.
Note that all machines listed above are being exclusively used for crunching. Under Windows they had a standard WinXP install and a standard set of tweaking to turn off anything (eg decorations, etc) that might impact performance. The only software installed was Cpuz, Everest and BOINC. Although I often use overclocking where appropriate, none of the above were overclocked at any point.
The distro being used for testing performance under Linux is PCLinuxOS TR4 which is installed from the liveCD. It runs with a KDE gui and I've not turned anything off. After installing from the liveCD I install BOINC and let it go.
EDIT:
I've now added two further machines to the above list. One is a 1.7GHz P4 Williamette and the other is an AMD Barton 2500+ which is running overclocked at 3200+ speeds (11x200 = 2.2GHz instead of the stock 11x166 = 1.833GHz). After a couple of hours of running (<10% crunched) the P4 is showing an apparent 8% decrease in CCH whilst the AMD is showing a 34% increase. My experience has been that the crunching tends to pick up speed a little in the middle stages after being a little slow in the early stages and again towards the end. I expect the P4 will gain a little and will finish about level pegging with the windows app whilst the AMD will also gain and probably finish with an increase of more than 40%.
So the upshot of all this is that if you are running any sort of PIII or an AMD you should use Linux at the moment if possible. This may all change when a new Windows app is released or when Akos gets to do his thing :).
If anyone is interested, the Linux distro I chose to use is PCLinuxOS 2007 TR-4. Check it out on Distrowatch. It's #3 on the Page Hit ranking and rising fast. Having installed it a few times now over the last couple of days I can really see why people like it :). It's really the first Linux distro I've ever used and it's very user friendly (for Linux/unix) :). I have the advantage of being familiar with BSD unixes and the unix CLI but the KDE gui of PCLOS is so easy for windows users to understand that this is a perfect way to introduce yourself to the joys of unix.
Wow, big differences. I think it's great you're making all this effort, Gary. Hopefully it will get us a bit closer to understanding what's going on.
In case you want to add the findings on my box to the list later:
1.6 GHz Core Duo T2060, 1 MB of L2 Cache per core... the core is called "Yonah", according to CPU-Z. Unless I messed up the calculations, there is a 40% increase between Windows XP and Ubuntu 7.04.
Maybe there's something left to tweak under Linux.
I think it would be instructive to include the absolute CCH value for each computer (and several more) in a table in the upcoming thread. After all we don't want to scare off Windows users, if you see that your CCH under windows for a particular CPU type is OK in comparison to Linux, there's no need to even think about switching OSes (at least not for E@H crunching).
Great idea. I might reattach my Venice under Windows for a single WU to get an estimate for that kind of CPU, but only if no one has already done that, and if it's relevant with AMD CPUs (or do we already agree that AMDs ALWAYS suck under Windows? not quite sure)
Great idea. I might reattach my Venice under Windows for a single WU to get an estimate for that kind of CPU, but only if no one has already done that, and if it's relevant with AMD CPUs (or do we already agree that AMDs ALWAYS suck under Windows? not quite sure)
Yep, I know it's not Windows
)
Yep, I know it's not Windows as such which is to blame. I dual booted before, with different projects or app versions, and never came across more than maybe 5% difference, one way or the other. I think some CPUs just react adversely to the current science app using the Windows code. Strange that your Pent M doesn't, though. Seeing the performance increase on both the P3 and the Core I thought it would. Yeah well, worst case one has to try it out for one's own CPU if there are no reliable rules about it. Still, I would appreciate some ideas from the project staff about the possible reasons. Bernd, are you reading this? Could you perhaps share some ideas with us curious volunteers? ;-)
Actually, I think I will. The new Ubuntu has some really nice programs on board, so I have no problems working under Linux, and most of the important hardware (graphics, ACPI, dual core support, WLAN, even the function keys*) works fine either out of the box or with some tweaking. This product series of Samsung notebooks seems to be suited well for Linux, especially Debian/Ubuntu, from what you read on the net. I'm also confident to get my IPSec up and running quite easily, or at least I'll know who to ask, so, no great problems there (I hope). My sound is annoying me a bit atm (meaning I don't have any), which seems to be a rather common bug with Feisty Fawn, but I'll force that to cooperate sooner or later.
So, it looks like there are not many arguments against running Linux as "primary" OS on this box. As you said, 30% more performance are a huge difference. It would mean I would finish a day's share of WUs in about 17 hours ;-) not bad, isn't it?
*Hope that is the correct English term. I mean when you press "Fn" and an F key and have features like switch off the sound or display or activate WLAN.
Hi Annika! Sounds like
)
Hi Annika!
Sounds like finally running Linux on a notebook is a real alternative now. I was never very convinced about WLAN support and things like power saving (dynamic CPU frequency switching).
Anyway...I'll try to crunch some more units under XP and wait and see what happens. We've seen strange things happening when you allow the graphics to display(slowdown even after closing the graphics again), it must be some equally strange effect that eats up CPU cycles on some systems, and not on others.
It would be cool if others who suspect a poor performance under this or that OS could try to run a few units on the same host but with a different OS (Live-CD DVD Linux distributions come in handy here).
Enjoy the rest of the holiday, too bad you have to study tomorrow (sorry for reminding you :-) :-) :-) :-) ).
CU
BRM
Actually, I'm studying now
)
Actually, I'm studying now ;-) as the code for tomorrow won't prepare itself and I had absolutely no time to do it earlier. Gotta get used to it I guess, the IT business is not exactly "nine to five" either...
Yeah, Linux on a laptop works okay I think. CT suspects about 5% less battery life, because power saving is not quite as efficient yet, but I think I can live with that atm. Still not figured out my sound problems, though...
RE: Still not figured out
)
Eh? You say something? ;-)
RE: Gotta get used to it I
)
So true!! But there's worse :-)
CU
BRM
RE: Surprisingly, on my
)
Sorry for not updating my progress but this difference between Linux and Windows on Intel hardware is turning into quite a big deal for me. I'm up to five separate installations now (and a sixth on the way) and all of them are showing a significant improvement for the 4.21 Linux app.
So far I've done 2 x Coppermine dual PIII 866 Xeons, 1 x Coppermine dual PIII 866, 1 x Tualatin dual PIII 1266 (the S version with double L2 cache), and 1 x Tualatin single PIII 1200 (standard L2 cache). The sixth install is a P4 Williamette 1.7GHz as I want to see how the early (and terrible!!) P4 architecture behaves.
Only one machine has actually had results returned and validated so far but all of the others are between 25% and 75% complete and this is enough to start making predictions. Assuming that the team have done the best job possible in matching allocated credit to the actual work performed, I think that credit per cpu hour (CCH) is a worthwhile yardstick to use for comparing crunching efficiency of the two operating systems. My own observation is that CCH is remarkably uniform for a given machine/OS combination and relatively independent of the actual result being crunched.
There is quite an architecture change between Coppermine PIIIs and Tualatin PIIIs which is why I was keen to test out some Tualatins. The Tualatin PIII had a quite short lifecycle as Intel killed it off to make way for the Williamette (and then Northwood) P4s. Once again Marketing won at the expense of good engineering :). The Tualatin was later resurrected and further improved to become the Pentium M.
As mentioned previously, I intend to write my findings into a separate thread but with the expansion of installs, that will certainly be a little later on :). In the meantime here is a summary of preliminary results - note C means Coppermine and T means Tualatin and the % increase is for the 4.21 Linux app as compared to the 4.17 windows app.
*PIII 866 Dual Xeon (C) - 37% increase in CCH (estimated - results @ 75% done)
*PIII 866 Dual (C) - 32% increase in CCH (estimated - results @ 25% done)
*PIII 1266 Dual (T) - 30% increase in CCH (estimated - results @ 40% done)
*PIII 1200 Single (T) - 28% increase in CCH (estimated - results @ 50% done)
Note that all machines listed above are being exclusively used for crunching. Under Windows they had a standard WinXP install and a standard set of tweaking to turn off anything (eg decorations, etc) that might impact performance. The only software installed was Cpuz, Everest and BOINC. Although I often use overclocking where appropriate, none of the above were overclocked at any point.
The distro being used for testing performance under Linux is PCLinuxOS TR4 which is installed from the liveCD. It runs with a KDE gui and I've not turned anything off. After installing from the liveCD I install BOINC and let it go.
EDIT:
I've now added two further machines to the above list. One is a 1.7GHz P4 Williamette and the other is an AMD Barton 2500+ which is running overclocked at 3200+ speeds (11x200 = 2.2GHz instead of the stock 11x166 = 1.833GHz). After a couple of hours of running (<10% crunched) the P4 is showing an apparent 8% decrease in CCH whilst the AMD is showing a 34% increase. My experience has been that the crunching tends to pick up speed a little in the middle stages after being a little slow in the early stages and again towards the end. I expect the P4 will gain a little and will finish about level pegging with the windows app whilst the AMD will also gain and probably finish with an increase of more than 40%.
So the upshot of all this is that if you are running any sort of PIII or an AMD you should use Linux at the moment if possible. This may all change when a new Windows app is released or when Akos gets to do his thing :).
If anyone is interested, the Linux distro I chose to use is PCLinuxOS 2007 TR-4. Check it out on Distrowatch. It's #3 on the Page Hit ranking and rising fast. Having installed it a few times now over the last couple of days I can really see why people like it :). It's really the first Linux distro I've ever used and it's very user friendly (for Linux/unix) :). I have the advantage of being familiar with BSD unixes and the unix CLI but the KDE gui of PCLOS is so easy for windows users to understand that this is a perfect way to introduce yourself to the joys of unix.
Cheers,
Gary.
Wow, big differences. I think
)
Wow, big differences. I think it's great you're making all this effort, Gary. Hopefully it will get us a bit closer to understanding what's going on.
In case you want to add the findings on my box to the list later:
1.6 GHz Core Duo T2060, 1 MB of L2 Cache per core... the core is called "Yonah", according to CPU-Z. Unless I messed up the calculations, there is a 40% increase between Windows XP and Ubuntu 7.04.
Yes, amazing! I'm glad my
)
Yes, amazing!
I'm glad my dual P III 866 Coppermine was running on Linux right from the start.
Funny I don't get a diff between my two Pentium M incarnations:
results under Win XP Home
results under Linux
Maybe there's something left to tweak under Linux.
I think it would be instructive to include the absolute CCH value for each computer (and several more) in a table in the upcoming thread. After all we don't want to scare off Windows users, if you see that your CCH under windows for a particular CPU type is OK in comparison to Linux, there's no need to even think about switching OSes (at least not for E@H crunching).
CU
BRM
Great idea. I might reattach
)
Great idea. I might reattach my Venice under Windows for a single WU to get an estimate for that kind of CPU, but only if no one has already done that, and if it's relevant with AMD CPUs (or do we already agree that AMDs ALWAYS suck under Windows? not quite sure)
RE: Great idea. I might
)
I am sure.
http://einsteinathome.org/node/192598&nowrap=true#66317
My VMWare Linux Cruncher(X2@2,63GHz):
http://einsteinathome.org/host/922868/tasks
http://einsteinathome.org/host/922404/tasks
My father's X2 5000+(2,6GHz/Win):
http://einsteinathome.org/host/902593/tasks
Running your Venice under Windows would be just waisting time.
cu,
Michael