Re: Benchmarking 2.2 and 2.4 using hdparm and dbench 1.1
[regarding the buffer cache hash size and bad performance on machines with little memory... (<32MB)] On Tue, 9 Jan 2001, Anton Blanchard wrote: > > Where is the size defined, and is it easy to modify? > > Look in fs/buffer.c:buffer_init() I experimented some, and increasing the huffer cache hash to the 2.2 levels helped a lot, especially for 16 MB memory. The difference is huge, 64 kB in 2.2 vs 1 kB in 2.4 for a 32 MB memory machine. > I havent done any testing on slow hardware and the high end stuff is > definitely performing better in 2.4, but I agree we shouldn't forget > about the slower stuff. Being able to tune the machine for both high and low end systems is neccessary, and if Linux can tune itself, that's of course the best. > Narrowing down where the problem is would help. My guess is it is a TCP > problem, can you check if it is performing worse in your case? (eg ftp > something against 2.2 and 2.4) Nope, TCP performance seems more or less unchanged. I will keep investigating, and get back when I have more info. /Tobias - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Re: Benchmarking 2.2 and 2.4 using hdparm and dbench 1.1
[regarding the buffer cache hash size and bad performance on machines with little memory... (32MB)] On Tue, 9 Jan 2001, Anton Blanchard wrote: Where is the size defined, and is it easy to modify? Look in fs/buffer.c:buffer_init() I experimented some, and increasing the huffer cache hash to the 2.2 levels helped a lot, especially for 16 MB memory. The difference is huge, 64 kB in 2.2 vs 1 kB in 2.4 for a 32 MB memory machine. I havent done any testing on slow hardware and the high end stuff is definitely performing better in 2.4, but I agree we shouldn't forget about the slower stuff. Being able to tune the machine for both high and low end systems is neccessary, and if Linux can tune itself, that's of course the best. Narrowing down where the problem is would help. My guess is it is a TCP problem, can you check if it is performing worse in your case? (eg ftp something against 2.2 and 2.4) Nope, TCP performance seems more or less unchanged. I will keep investigating, and get back when I have more info. /Tobias - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Re: Benchmarking 2.2 and 2.4 using hdparm and dbench 1.1
On Tuesday 09 January 2001 12:08, Anton Blanchard wrote: > > Where is the size defined, and is it easy to modify? > > Look in fs/buffer.c:buffer_init() > > > I noticed that /proc/sys/vm/freepages is not writable any more. Is there > > any reason for this? > > I am not sure why. > It can probably be made writeable, within limits (caused by zones...) But the interesting part is that 2.4 tries to estimate how much memory it will need shortly (inactive_target) and try to keep that amount inactive clean (inactive_clean) - clean inactive memory can be freed and reused very quickly. cat /proc/meminfo My feeling is that, for now, keeping it untuneable can help us in finding fixable cases... /RogerL -- Home page: http://www.norran.net/nra02596/ - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Re: Benchmarking 2.2 and 2.4 using hdparm and dbench 1.1
> Where is the size defined, and is it easy to modify? Look in fs/buffer.c:buffer_init() > I noticed that /proc/sys/vm/freepages is not writable any more. Is there > any reason for this? I am not sure why. > Hmm... I'm still using samba 2.0.7. I'll try 2.2 to see if it > helps. What are tdb spinlocks? samba 2.2 uses tdb which is an SMP safe gdbm like database. By default it uses byte range fcntl locks to provide locking, but has the option of using spinlocks (./configure --with-spinlocks). I doubt it would make a difference on your setup. > Have you actually compared the same setup with 2.2 and 2.4 kernels and a > single client transferring a large file, preferably from a slow server > with little memory? Most samba servers that people benchmark are fast > computers with lots of memory. So far, every major kernel upgrade has > given me a performance boost, even for slow computers, and I would hate to > see that trend break for 2.4... I havent done any testing on slow hardware and the high end stuff is definitely performing better in 2.4, but I agree we shouldn't forget about the slower stuff. Narrowing down where the problem is would help. My guess is it is a TCP problem, can you check if it is performing worse in your case? (eg ftp something against 2.2 and 2.4) Anton - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Re: Benchmarking 2.2 and 2.4 using hdparm and dbench 1.1
On Tuesday 09 January 2001 12:08, Anton Blanchard wrote: Where is the size defined, and is it easy to modify? Look in fs/buffer.c:buffer_init() I noticed that /proc/sys/vm/freepages is not writable any more. Is there any reason for this? I am not sure why. It can probably be made writeable, within limits (caused by zones...) But the interesting part is that 2.4 tries to estimate how much memory it will need shortly (inactive_target) and try to keep that amount inactive clean (inactive_clean) - clean inactive memory can be freed and reused very quickly. cat /proc/meminfo My feeling is that, for now, keeping it untuneable can help us in finding fixable cases... /RogerL -- Home page: http://www.norran.net/nra02596/ - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Re: Benchmarking 2.2 and 2.4 using hdparm and dbench 1.1
On Fri, 5 Jan 2001, Anton Blanchard wrote: > > > 1) Why does the hdbench numbers go down for 2.4 (only) when 32 MB is used? > >I fail to see how that matters, especially for the '-T' test. > > When I did some tests long ago, hdparm was hitting the buffer cache hash > table pretty hard in 2.4 compared to 2.2 because it is now smaller. However > as davem pointed out, most things don't do such things so resizing the hash > table just for this is a waste. Where is the size defined, and is it easy to modify? > Since the hash is based on RAM, it may end up being big enough on the 128M > machine. Maybe. I have been experimenting some more, and I see that the less memory i have, kswapd takes more and more CPU (more than 10% for some cases) when I am doing a continuous read from a block device. I noticed that /proc/sys/vm/freepages is not writable any more. Is there any reason for this? > > The reason for doing the benchmarks in the first place is that my 32MB P90 > > at home really does perform noticeably worse with samba using 2.4 kernels > > than using 2.2 kernels, and that bugs me. I have no hard numbers for that > > machine (yet). If they will show anything extra, I will post them here. > > What exactly are you seeing? I first noticed the slowdown because the load meter LEDs on my ethernet hub did not go as high with 2.4 as it did with 2.2. A simple test, transferring a large file using smbclient, did in deed show a decrease in performance, both for a localhost and a remote file transfer. This in spite of the tcp transfer rate beeing (much) higher in 2.4 than in 2.2. > > Btw, has anyone else noticed samba slowdowns when going from 2.2 to 2.4? > > I am seeing good results with 2.4 + samba 2.2 using tdb spinlocks. Hmm... I'm still using samba 2.0.7. I'll try 2.2 to see if it helps. What are tdb spinlocks? Have you actually compared the same setup with 2.2 and 2.4 kernels and a single client transferring a large file, preferably from a slow server with little memory? Most samba servers that people benchmark are fast computers with lots of memory. So far, every major kernel upgrade has given me a performance boost, even for slow computers, and I would hate to see that trend break for 2.4... /Tobias - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Re: Benchmarking 2.2 and 2.4 using hdparm and dbench 1.1
> 1) Why does the hdbench numbers go down for 2.4 (only) when 32 MB is used? >I fail to see how that matters, especially for the '-T' test. When I did some tests long ago, hdparm was hitting the buffer cache hash table pretty hard in 2.4 compared to 2.2 because it is now smaller. However as davem pointed out, most things don't do such things so resizing the hash table just for this is a waste. Since the hash is based on RAM, it may end up being big enough on the 128M machine. > 3) The 2.2 kernels outperform the 2.4 kernels for few clients (see >especially the "dbench 1" numbers for the PII-128M. Oops! dbench was sleeping up to 1 second between starting the clock and starting the benchmark. This made small benchmarks (ie especially dbench 1) somewhat variable. Grab the latest version from pserver.samba.org (if you havent already). > The reason for doing the benchmarks in the first place is that my 32MB P90 > at home really does perform noticeably worse with samba using 2.4 kernels > than using 2.2 kernels, and that bugs me. I have no hard numbers for that > machine (yet). If they will show anything extra, I will post them here. What exactly are you seeing? > Btw, has anyone else noticed samba slowdowns when going from 2.2 to 2.4? I am seeing good results with 2.4 + samba 2.2 using tdb spinlocks. Anton - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Re: Benchmarking 2.2 and 2.4 using hdparm and dbench 1.1
1) Why does the hdbench numbers go down for 2.4 (only) when 32 MB is used? I fail to see how that matters, especially for the '-T' test. When I did some tests long ago, hdparm was hitting the buffer cache hash table pretty hard in 2.4 compared to 2.2 because it is now smaller. However as davem pointed out, most things don't do such things so resizing the hash table just for this is a waste. Since the hash is based on RAM, it may end up being big enough on the 128M machine. 3) The 2.2 kernels outperform the 2.4 kernels for few clients (see especially the "dbench 1" numbers for the PII-128M. Oops! dbench was sleeping up to 1 second between starting the clock and starting the benchmark. This made small benchmarks (ie especially dbench 1) somewhat variable. Grab the latest version from pserver.samba.org (if you havent already). The reason for doing the benchmarks in the first place is that my 32MB P90 at home really does perform noticeably worse with samba using 2.4 kernels than using 2.2 kernels, and that bugs me. I have no hard numbers for that machine (yet). If they will show anything extra, I will post them here. What exactly are you seeing? Btw, has anyone else noticed samba slowdowns when going from 2.2 to 2.4? I am seeing good results with 2.4 + samba 2.2 using tdb spinlocks. Anton - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Re: Benchmarking 2.2 and 2.4 using hdparm and dbench 1.1
On Fri, 5 Jan 2001, Anton Blanchard wrote: 1) Why does the hdbench numbers go down for 2.4 (only) when 32 MB is used? I fail to see how that matters, especially for the '-T' test. When I did some tests long ago, hdparm was hitting the buffer cache hash table pretty hard in 2.4 compared to 2.2 because it is now smaller. However as davem pointed out, most things don't do such things so resizing the hash table just for this is a waste. Where is the size defined, and is it easy to modify? Since the hash is based on RAM, it may end up being big enough on the 128M machine. Maybe. I have been experimenting some more, and I see that the less memory i have, kswapd takes more and more CPU (more than 10% for some cases) when I am doing a continuous read from a block device. I noticed that /proc/sys/vm/freepages is not writable any more. Is there any reason for this? The reason for doing the benchmarks in the first place is that my 32MB P90 at home really does perform noticeably worse with samba using 2.4 kernels than using 2.2 kernels, and that bugs me. I have no hard numbers for that machine (yet). If they will show anything extra, I will post them here. What exactly are you seeing? I first noticed the slowdown because the load meter LEDs on my ethernet hub did not go as high with 2.4 as it did with 2.2. A simple test, transferring a large file using smbclient, did in deed show a decrease in performance, both for a localhost and a remote file transfer. This in spite of the tcp transfer rate beeing (much) higher in 2.4 than in 2.2. Btw, has anyone else noticed samba slowdowns when going from 2.2 to 2.4? I am seeing good results with 2.4 + samba 2.2 using tdb spinlocks. Hmm... I'm still using samba 2.0.7. I'll try 2.2 to see if it helps. What are tdb spinlocks? Have you actually compared the same setup with 2.2 and 2.4 kernels and a single client transferring a large file, preferably from a slow server with little memory? Most samba servers that people benchmark are fast computers with lots of memory. So far, every major kernel upgrade has given me a performance boost, even for slow computers, and I would hate to see that trend break for 2.4... /Tobias - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Re: Benchmarking 2.2 and 2.4 using hdparm and dbench 1.1
On Wed, 3 Jan 2001, Daniel Phillips wrote: > Tobias Ringstrom wrote: > > 3) The 2.2 kernels outperform the 2.4 kernels for few clients (see > >especially the "dbench 1" numbers for the PII-128M. Oops! > > I noticed that too. Furthermore I noticed that the results of the more > heavily loaded tests on the whole 2.4.0 series tend to be highly > variable (usually worse) if you started by moving the whole disk through > cache, e.g., fsck on a damaged filesystem. Yes, they do seem to vary a lot. > It would be great if you could track the ongoing progress - you could go > so far as to automatically download the latest patch and rerun the > tests. (We have a script like that here to keep our lxr/cvs tree > current.) And yes, it gets more important to consider some of the other > usage patterns so we don't end up with self-fullfilling prophecies. I was thinking about an automatic test, build, modify lilo, reboot cycle for a while, but I don't think it's worth it. Benchmarking is hard, and making it automatic is probably even harder, not mentioning trying to interpret the numbers... Probably "Samba feels slower" works quite well. :-) But then it is even unclear to me what the vm people are trying to optimize for. Probably a system that "feels good", which according to myself above, may actually be a good criteria, although a but imprecise. Oh, well... > For benchmarking it would be really nice to have a way of emptying > cache, beyond just syncing. I took a look at that last week and > unfortunately it's not trivial. The things that have to be touched are > optimized for the steady-state running case and tend to take their > marching orders from global variables and embedded heuristics that you > don't want to mess with. Maybe I'm just looking at this problem the > wrong way because the shortest piece of code I can imagine for doing > this would be 1-200 lines long and would replicate a lot of the > functionality of page_launder and flush_dirty_pages, in other words it > would be a pain to maintain. How about allocating lots of memory and locking it in memory? I have not looked at the source, but it seems (using strace) that hdbench uses shm to do just that. I'll dig into the hdbench code and try to make a program that empties the cache. /Tobias - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Re: Benchmarking 2.2 and 2.4 using hdparm and dbench 1.1
Tobias Ringstrom wrote: > 3) The 2.2 kernels outperform the 2.4 kernels for few clients (see >especially the "dbench 1" numbers for the PII-128M. Oops! I noticed that too. Furthermore I noticed that the results of the more heavily loaded tests on the whole 2.4.0 series tend to be highly variable (usually worse) if you started by moving the whole disk through cache, e.g., fsck on a damaged filesystem. > The reason for doing the benchmarks in the first place is that my 32MB P90 > at home really does perform noticeably worse with samba using 2.4 kernels > than using 2.2 kernels, and that bugs me. I have no hard numbers for that > machine (yet). If they will show anything extra, I will post them here. > Btw, has anyone else noticed samba slowdowns when going from 2.2 to 2.4? Again, yes, I saw that. > Wow! You made it all the way down here. Congratulations! :-) Heh. Yes, then I read it all again backwards. I'll respectfully bow out of the benchmarking business now. :-) It would be great if you could track the ongoing progress - you could go so far as to automatically download the latest patch and rerun the tests. (We have a script like that here to keep our lxr/cvs tree current.) And yes, it gets more important to consider some of the other usage patterns so we don't end up with self-fullfilling prophecies. For benchmarking it would be really nice to have a way of emptying cache, beyond just syncing. I took a look at that last week and unfortunately it's not trivial. The things that have to be touched are optimized for the steady-state running case and tend to take their marching orders from global variables and embedded heuristics that you don't want to mess with. Maybe I'm just looking at this problem the wrong way because the shortest piece of code I can imagine for doing this would be 1-200 lines long and would replicate a lot of the functionality of page_launder and flush_dirty_pages, in other words it would be a pain to maintain. -- Daniel - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Benchmarking 2.2 and 2.4 using hdparm and dbench 1.1
I have been torturing a couple of boxes and came up with these benchmark results. I have also enclosed the script used to do the benchmark, and I am well aware that this is a very specialized benchmark, testing only limited parts of the kernel, and so on, BUT I am convinced that I'm seeing something interesting things in the results anyway. :-) There are numbers for two different machines, and each test has been run five times in succession, with a sync and a sleep in between. The kernels were optimized for Pentium Classic, and only contained IDE support (e.g. no SCSI, networking, etc). The only processes running besides the benchmark script were kernel daemons and getty:s. Both machines had plenty of swap space available. There are some numbers that I think are interesting: 1) Why does the hdbench numbers go down for 2.4 (only) when 32 MB is used? I fail to see how that matters, especially for the '-T' test. 2) The 2.4 kernels outperform the 2.2 kernels for many clients (see the "dbench 10" numbers). This matters a lot. Great! 3) The 2.2 kernels outperform the 2.4 kernels for few clients (see especially the "dbench 1" numbers for the PII-128M. Oops! The reason for doing the benchmarks in the first place is that my 32MB P90 at home really does perform noticeably worse with samba using 2.4 kernels than using 2.2 kernels, and that bugs me. I have no hard numbers for that machine (yet). If they will show anything extra, I will post them here. Btw, has anyone else noticed samba slowdowns when going from 2.2 to 2.4? Anyway, any help explaining/fixing points 1 and 3 would be highly appreciated! /Tobias [All numbers in MB/s, bigger are better.] == 133 MHz Pentium, DMA, 80M 2.2.16-22: hdparm -T 37.43 37.43 37.32 37.32 37.43 hdparm -t 6.086.146.156.106.08 dbench 110.4302 10.0796 10.2559 10.2258 13.5464 dbench 57.97753 8.13792 7.66108 7.8526 7.44329 dbench 10 5.78309 5.58762 5.76388 5.54761 5.94415 2.2.18: hdparm -T 37.87 37.87 37.65 37.54 37.65 hdparm -t 6.046.046.106.116.07 dbench 19.98084 9.19558 10.1023 9.78034 10.7593 dbench 57.5335 7.85761 7.90051 8.19119 7.78873 dbench 10 5.98423 5.99556 5.84676 6.02366 5.87104 2.2.19pre3: hdparm -T 37.98 37.43 37.21 37.43 37.87 hdparm -t 6.086.086.066.116.09 dbench 110.2117 11.2996 12.017 11.3003 12.1677 dbench 56.51203 6.80555 6.46566 6.66772 6.56693 dbench 10 5.55781 5.68997 5.43493 5.58688 5.32528 2.4.0-prerelease: hdparm -T 38.21 38.21 38.32 38.21 38.21 hdparm -t 6.076.246.146.306.26 dbench 14.23029 3.89185 10.27 14.2546 14.2719 dbench 58.63648 9.21302 11.0506 9.64396 10.8724 dbench 10 7.12402 6.92772 8.02011 7.65119 8.12557 == 133 MHz Pentium, DMA, 32M 2.2.16-22: hdparm -T 37.10 37.76 37.98 37.54 37.32 hdparm -t 6.016.046.075.936.06 dbench 110.0048 8.59813 8.49598 9.59796 9.04642 dbench 53.99638 4.59673 4.08813 3.90389 4.30821 dbench 10 2.6141 2.69585 2.63201 2.59543 2.61565 2.2.18: hdparm -T 37.21 37.98 37.98 37.98 37.76 hdparm -t 5.945.986.036.056.03 dbench 19.66983 10.3501 9.77682 9.44597 10.0551 dbench 55.32253 5.31517 5.9542 5.83028 5.22383 dbench 10 2.86753 2.85903 2.83746 2.93299 2.84175 2.2.19pre3: hdparm -T 36.99 38.21 37.98 37.87 37.10 hdparm -t 6.086.156.146.146.14 dbench 18.64248 7.73716 7.42123 7.6462 9.58718 dbench 54.57973 4.5689 4.32196 4.50044 4.68734 dbench 10 2.56748 2.55824 2.56042 2.54797 2.59391 2.4.0-prerelease: hdparm -T 37.32 37.21 37.32 37.32 37.21 hdparm -t 5.875.705.345.095.07 dbench 14.41358 7.43094 8.76442 8.64346 9.45806 dbench 54.42467 4.22284 4.89167 4.48322 5.08206 dbench 10 4.19795 5.6161 4.09045 4.09799 4.55435 == 400 MHz Mobile Pentium II, UDMA(33), mem=128M 2.2.18: hdparm -T 82.05 82.05 82.05 82.05 82.05 hdparm -t 9.859.8210.09 10.08 9.83 dbench 145.8926 48.1175 47.7244 47.7418 47.457 dbench 518.4021 20.7948 22.3345 20.4869 14.2627 dbench 10 13.1219 10.0275 10.8189 13.7482 9.51982 2.2.19pre3: hdparm -T 81.53 82.05 82.05 82.05 82.05 hdparm -t 9.889.899.869.889.85 dbench 138.752 43.6587 43.1486 41.004 42.4516 dbench 526.6624 27.4806 26.1421 28.9142 24.6634 dbench 10 14.9424 14.6862 14.2302 14.6954 14.8282 2.4.0-prerelease: hdparm -T 82.05 83.12 83.12 83.12 83.12 hdparm -t 10.08 10.08
Re: Benchmarking 2.2 and 2.4 using hdparm and dbench 1.1
Tobias Ringstrom wrote: 3) The 2.2 kernels outperform the 2.4 kernels for few clients (see especially the "dbench 1" numbers for the PII-128M. Oops! I noticed that too. Furthermore I noticed that the results of the more heavily loaded tests on the whole 2.4.0 series tend to be highly variable (usually worse) if you started by moving the whole disk through cache, e.g., fsck on a damaged filesystem. The reason for doing the benchmarks in the first place is that my 32MB P90 at home really does perform noticeably worse with samba using 2.4 kernels than using 2.2 kernels, and that bugs me. I have no hard numbers for that machine (yet). If they will show anything extra, I will post them here. Btw, has anyone else noticed samba slowdowns when going from 2.2 to 2.4? Again, yes, I saw that. Wow! You made it all the way down here. Congratulations! :-) Heh. Yes, then I read it all again backwards. I'll respectfully bow out of the benchmarking business now. :-) It would be great if you could track the ongoing progress - you could go so far as to automatically download the latest patch and rerun the tests. (We have a script like that here to keep our lxr/cvs tree current.) And yes, it gets more important to consider some of the other usage patterns so we don't end up with self-fullfilling prophecies. For benchmarking it would be really nice to have a way of emptying cache, beyond just syncing. I took a look at that last week and unfortunately it's not trivial. The things that have to be touched are optimized for the steady-state running case and tend to take their marching orders from global variables and embedded heuristics that you don't want to mess with. Maybe I'm just looking at this problem the wrong way because the shortest piece of code I can imagine for doing this would be 1-200 lines long and would replicate a lot of the functionality of page_launder and flush_dirty_pages, in other words it would be a pain to maintain. -- Daniel - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Re: Benchmarking 2.2 and 2.4 using hdparm and dbench 1.1
On Wed, 3 Jan 2001, Daniel Phillips wrote: Tobias Ringstrom wrote: 3) The 2.2 kernels outperform the 2.4 kernels for few clients (see especially the "dbench 1" numbers for the PII-128M. Oops! I noticed that too. Furthermore I noticed that the results of the more heavily loaded tests on the whole 2.4.0 series tend to be highly variable (usually worse) if you started by moving the whole disk through cache, e.g., fsck on a damaged filesystem. Yes, they do seem to vary a lot. It would be great if you could track the ongoing progress - you could go so far as to automatically download the latest patch and rerun the tests. (We have a script like that here to keep our lxr/cvs tree current.) And yes, it gets more important to consider some of the other usage patterns so we don't end up with self-fullfilling prophecies. I was thinking about an automatic test, build, modify lilo, reboot cycle for a while, but I don't think it's worth it. Benchmarking is hard, and making it automatic is probably even harder, not mentioning trying to interpret the numbers... Probably "Samba feels slower" works quite well. :-) But then it is even unclear to me what the vm people are trying to optimize for. Probably a system that "feels good", which according to myself above, may actually be a good criteria, although a but imprecise. Oh, well... For benchmarking it would be really nice to have a way of emptying cache, beyond just syncing. I took a look at that last week and unfortunately it's not trivial. The things that have to be touched are optimized for the steady-state running case and tend to take their marching orders from global variables and embedded heuristics that you don't want to mess with. Maybe I'm just looking at this problem the wrong way because the shortest piece of code I can imagine for doing this would be 1-200 lines long and would replicate a lot of the functionality of page_launder and flush_dirty_pages, in other words it would be a pain to maintain. How about allocating lots of memory and locking it in memory? I have not looked at the source, but it seems (using strace) that hdbench uses shm to do just that. I'll dig into the hdbench code and try to make a program that empties the cache. /Tobias - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/