Joerg Schilling schrieb: > Thomas Maier-Komor <[EMAIL PROTECTED]> wrote: > > > Let me add that if I run > > bzip2 -d < linux-2.6.17.3.tar.bz2 | sdd -onull bs=10k -t > > I get ~ 14 MB/z read speed in sdd >
OK, you have a faster CPU. But I don't really see what we can make from this info. mtwrite.so will create massive amounts of nonsequential accesses that must be resorted by the OS (i.e. the disk drivers strategy routine) into sequential operations. If this cannot be performed you get more seeks. So this is the basic idea of using mtwrite.so on top of ZFS. ZFS translates random writes into sequential disk accesses. >> I know that star/tar's performance regress on UFS filesystems if you preload >> mtwrite.so. I performed my tests on a ZFS filesystem. I don't have a recent >> linux kernel at hand, because it has been some time since I last used it. >> But all Linux are equal, as you know, and here are some tests with an older >> linux kernel archive: > > But why is my test 6.7 times slower when using your mtwrite.so? > I have no idea. What Solaris release are you running? > >> [EMAIL PROTECTED]:/var/a$ star --version >> star: star 1.5a57 (sparc-sun-solaris2.8) >> >> Copyright (C) 1985, 88-90, 92-96, 98, 99, 2000-2005 Jörg Schilling >> This is free software; see the source for copying conditions. There is NO >> warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. > > A bit old... the recent version is 1.5a77-pre, your version is 2 years old. > Were there any changes that could result in performance differences? I could update to more recent version - but I've only a 64k dialup, so this is always a hassel when packages are dependent on a bunch of things. I will check if I can do it... > [snipped results of star with option -no-fsync] > > so the best star standalone extract on UFS is 11x faster than the one with > mtwrite.so > OK, -no-fsync makes mtwrite.so pretty superfluous. Then star can take advantage of several caching mechanisms of the OS. Keep fsync on and then let's see what benefits mtwrite.so can offer, as it then enables star to have several outstanding fsync requests in parallel. >> >> Be aware that the attached filesystems are hosten on fast SCSI disks. The >> UFS consists of a mirror of two 15k RPM 73G disks. The ZFS is located on a >> mirror of two 10k RPM 146G disks. Maybe this might be the reason that my >> tests perform much better in threaded mode... > > Why is your extract speed with mtwrite.so still ~ 4x better then mine? > I don't really know? Why is your's 4x worse? There are several differences between our two systems (Solaris release, host system architecture and bus architecture, storage architecture, processor and memory speed, ...). Maybe you could tell us some of these details to get a more concrete idea. Or maybe you could run a dtrace and show what is causing the degredation. You could try changing the amount of parallelity mtwrite.so pushes on your system, by setting MTWRITE_MAXTHREADS to some different value. The default 16 could be a little bit demanding for a PC. Thomas _______________________________________________ opensolaris-discuss mailing list [email protected]
