| |||||||||||||
Testing on old hardware is flawedTesting on old hardware is flawedPosted Aug 26, 2004 9:40 UTC (Thu) by hensema (guest, #980)Parent article: Looking at reiser4
Please don't test filesystems on old hardware. Current hardware has totally different cpu-speed/disk-speed ratios, while both the speed of the cpu and disk have risen tremendiously.
Reiserfs 4 has been written for modern hardware. It has no place on old hardware. And since there are no conversion tools available, there is no viable upgrade path anyway.
It would be really interesting to benchmark again on a > 2 Ghz machine with a disk which does > 40 MB/sec, in at least 512 MB of ram. Such a machine would be at the low end where reiser4 installations will occur (not because reiser4 won't work on older hardware, but just because most installations will occur on new hardware).
One of reiser4's problems on older machines is the high cpu usage. On modern hardware this problem will be less visible, since cpu power has outgrown disk speed by an order of magnutude.
(Log in to post comments)
Testing on old hardware is flawed Posted Aug 26, 2004 11:35 UTC (Thu) by Fats (subscriber, #14882) [Link] What you are actually saying is that reiserfs is more optimized for benchmarks then for real live. In real live you want to have the fastest disk access possible with the least CPU usage. For example on mail servers (with virus filters) you don't want the filesystem take away precious CPU time from the main task.
Testing on old hardware is flawed Posted Aug 26, 2004 11:47 UTC (Thu) by hensema (guest, #980) [Link] No, what I'm saying is that a disk is dog slow compared to a CPU. It pays to invest some extra CPU cycles to prevent unnescesary disk seeks. Packing data more tightly on disk helps too.
On a slow CPU with a relatively fast disk this doesn't work. Here you can affort to do some extra seeks or waste some space on disk. Therefore, reiser4 is optimized for current and future systems and won't give optimal performance on older systems like the one used for the little lwn.net benchmark.
Also, CPU usage is generally part of a good benchmark. It's even part of really bad ones, like untarring a kernel :-)
Testing on old hardware is flawed Posted Aug 26, 2004 17:12 UTC (Thu) by NAR (subscriber, #1313) [Link] I'm not sure that the CPU speed/disk speed ratio differs that much between the lwn.net test machine and the test machine you asked for. From your reasing I'd think reiser4 should be faster on old machines because there the disk is really slow. However, I might be wrong.
Testing on old hardware is flawed Posted Aug 26, 2004 21:22 UTC (Thu) by hensema (guest, #980) [Link] No, on old machines the disk is actually relatively fast. In absolute figures the disk is slow, maybe 10 MB/sec or slightly less. But current processors are more than 10-20 times faster while disks are only 4-5 times faster. And that's measured by throughput. Seeks are barely any faster than 4 years ago (they are mostly limited by the rotational speed of the platters).
Testing on old hardware is flawed Posted Sep 3, 2004 10:19 UTC (Fri) by gc (guest, #24112) [Link] Recent CPUs have very high frequencies but the counterpart is that they need more clock cycles to perform the same instructions. If you compare a P4-3GHz with a P3-300MHz, the clock ticks needed to execute a typical series of instructions is 3 to 4 times larger, hence the CPU power is overally "only" 3 times higher. I think your figures concerning the increase of CPU power are wrong.
Testing on old hardware is flawed Posted Aug 26, 2004 11:47 UTC (Thu) by bpearlmutter (subscriber, #14693) [Link] No, that is not what he is saying at all.Sometimes there is a tradeoff between thinking about where and how to put stuff on disk (CPU) and actual disk access. On a system like a mail server, you might expect the disk to be the bottleneck, with well below 100% CPU usage---at least, during periods when data is actually being moved on/off disk. This is precisely where, as the CPU/disk speed disparity grows, it makes sense to spend more CPU to optimise (reduce) actual disk access. I don't know if reiserfs4 succeeds in this, but if it does it would be of particular advantage on something like a mail/news server. And, it would not be properly benchmarked if the CPU/disk speed disparity were unrealistically low. Barak A. Pearlmutter <barak@cs.may.ie> Hamilton Institute, NUI Maynooth, Co. Kildare, Ireland http://www-bcl.cs.may.ie/~barak/
Testing on old hardware is flawed Posted Aug 27, 2004 14:53 UTC (Fri) by jeremiah (subscriber, #1221) [Link] On our DB server, we have lots of extra CPU cycles to burn and our Drives are constantly maxed out. I would gladly give up 25% - 50% of my CPU if it ment I got that much or more of an I/O imporovement from my filesystem.
I think this is just one of those place where the server side of Linux differs from the Desktop side of linux. Or more clearly (I hope ) is that the Mail server differs from the DB server side of things. RFS4 seems to be addressing high disk usage servers.
|
|||||||||||||