Lorenzo Milesi
2008-Dec-02 10:59 UTC
[Ocfs2-users] Please urgent help required - OCFS2 and VPN again
Hi all... I already wrote before on the list about the solution I have at a customer running DRBD8+OCFS2 on two remote sites connected via VPN. The different suggestions helped improving the situation, but still we're having big troubles. We've also upgraded the old server with a new and much more powerful one but there was nearly no improvement at all! The situation is resumed as: SITE A: Dual Core 2GHz Pentium, 1Gb ram, 1 SATA hdd for /, 3 SATA hdd in software raid5, DRBD on /dev/md0. SITE B: Quad Core 2.4HGz Pentium, 2Gb ram, 3 SATA HDD in software raid5, DRBD on /dev/md1. The two sites are connected using two ADSL, with TWO bonded VPN. Both machines run Debian Etch fully updated, kernel 2.6.26-bpo.1-686 SMP with deadline scheduler, DRBD 8.0.13, OCFS2 1.4.1-1. The shared data partition is 187G, 30 of which used. The recent upgrade to OCFS2 1.4 and kernel 2.6.26 didn't improve the performances as much as I expected. The main problems we have are: 1. very high load average: this was previously caused by very high iowait percentages, but with the new server the load is high while top says the machine is 99-100% idle! 2. very slow dir browsing: Sunil pointed me to the user guide, where he talks about inode stat. How can I raise inode cache memory? I've done several searches without result... The server actually uses less than 300Mb of ram out of the 1Gb installed... 3. very long umount time: I often (not always) experience an extremely long umount time. During the period while the process is executing iftop says there's a high usage of network transfer. I suppose it's transfering file locks, but is it possible that stays stuck for more than one hour, and still going? This is the configuration file of OCFS2. The quad-core is file-server-2. #/etc/ocfs2/cluster.conf node: ip_port = 7777 ip_address = 192.168.0.1 number = 0 name = file-server-1 cluster = ocfs2 node: ip_port = 7777 ip_address = 192.168.2.31 number = 1 name = file-server-2 cluster = ocfs2 cluster: node_count = 2 name = ocfs2 What is stunning me is that on file-server-2 we run a rsync backup during the night on a local machine on the network, and it takes less than 20m! Doing the same on the other server throws the load average to the stars! We're in a critical situation because this solution is deployed since a long time and it's not yet working as expected. If nobody has suggestion we have no problem in paying qualified support for solving these problems. In this case please contact me directly. Sunil, can I get Oracle support for this? Thank you. -- Lorenzo Milesi - lorenzo.milesi at yetopen.it YetOpen S.r.l. - http://www.yetopen.it/ C.so E. Filiberto, 74 23900 Lecco - ITALY - Tel 0341 220 205 - Fax 178 607 8199 GPG/PGP Key-Id: 0xE704E230 - http://keyserver.linux.it -------- D.Lgs. 196/2003 -------- Si avverte che tutte le informazioni contenute in questo messaggio sono riservate ed a uso esclusivo del destinatario. Nel caso in cui questo messaggio Le fosse pervenuto per errore, La invitiamo ad eliminarlo senza copiarlo, a non inoltrarlo a terzi e ad avvertirci non appena possibile. Grazie.
Luis Freitas
2008-Dec-02 16:48 UTC
[Ocfs2-users] Please urgent help required - OCFS2 and VPN again
Lorenzo, ?? My 2 cents. This is purely speculation, since I never worked with a environment like what you have there. ?? You have a very different configuration from what is usual with OCFS2. The filesystem is tested on systems that have a fast network connection between nodes, so it is probably not tuned to environments where the network bandwith is low. ??? You might get some improvement if you change some of the VM tunables (/proc/sys/vm/*). On 2.6 there are not much of them for the filesystem cache, and some seem to have no effect. ?? vfs_cache_pressure could give you some control on the quantity of inodes cached by the kernel. Try either increasing or decreasing. Some of the problems you relate, like the long time for umount might actually be caused by keeping a large amount of structures on memory, and since the network is slow a long time is needed to clear all of them. ?? Swappiness controls how aggressivelly pages are swapped out. Since you dont have swap on the OCFS2 filesystem it should not have much impact. (You dont have swap there, right?) But you may be able to force the kernel to release memory so that it can be used by the OCFS2. Again this could actually cause the problem to become worse. ?? There used to be a parameter that controls how much of the cache is used for inodes, and how much of it is used for data blocks, dcache_priorty.? But it no longer available on 2.6? and I could not find an equivalent. Regards, Luis --- On Tue, 12/2/08, Lorenzo Milesi <lorenzo.milesi at yetopen.it> wrote: From: Lorenzo Milesi <lorenzo.milesi at yetopen.it> Subject: [Ocfs2-users] Please urgent help required - OCFS2 and VPN again To: ocfs2-users at oss.oracle.com Date: Tuesday, December 2, 2008, 8:59 AM Hi all... I already wrote before on the list about the solution I have at a customer running DRBD8+OCFS2 on two remote sites connected via VPN. The different suggestions helped improving the situation, but still we're having big troubles. We've also upgraded the old server with a new and much more powerful one but there was nearly no improvement at all! The situation is resumed as: SITE A: Dual Core 2GHz Pentium, 1Gb ram, 1 SATA hdd for /, 3 SATA hdd in software raid5, DRBD on /dev/md0. SITE B: Quad Core 2.4HGz Pentium, 2Gb ram, 3 SATA HDD in software raid5, DRBD on /dev/md1. The two sites are connected using two ADSL, with TWO bonded VPN. Both machines run Debian Etch fully updated, kernel 2.6.26-bpo.1-686 SMP with deadline scheduler, DRBD 8.0.13, OCFS2 1.4.1-1. The shared data partition is 187G, 30 of which used. The recent upgrade to OCFS2 1.4 and kernel 2.6.26 didn't improve the performances as much as I expected. The main problems we have are: 1. very high load average: this was previously caused by very high iowait percentages, but with the new server the load is high while top says the machine is 99-100% idle! 2. very slow dir browsing: Sunil pointed me to the user guide, where he talks about inode stat. How can I raise inode cache memory? I've done several searches without result... The server actually uses less than 300Mb of ram out of the 1Gb installed... 3. very long umount time: I often (not always) experience an extremely long umount time. During the period while the process is executing iftop says there's a high usage of network transfer. I suppose it's transfering file locks, but is it possible that stays stuck for more than one hour, and still going? This is the configuration file of OCFS2. The quad-core is file-server-2. #/etc/ocfs2/cluster.conf node: ip_port = 7777 ip_address = 192.168.0.1 number = 0 name = file-server-1 cluster = ocfs2 node: ip_port = 7777 ip_address = 192.168.2.31 number = 1 name = file-server-2 cluster = ocfs2 cluster: node_count = 2 name = ocfs2 What is stunning me is that on file-server-2 we run a rsync backup during the night on a local machine on the network, and it takes less than 20m! Doing the same on the other server throws the load average to the stars! We're in a critical situation because this solution is deployed since a long time and it's not yet working as expected. If nobody has suggestion we have no problem in paying qualified support for solving these problems. In this case please contact me directly. Sunil, can I get Oracle support for this? Thank you. -- Lorenzo Milesi - lorenzo.milesi at yetopen.it YetOpen S.r.l. - http://www.yetopen.it/ C.so E. Filiberto, 74 23900 Lecco - ITALY - Tel 0341 220 205 - Fax 178 607 8199 GPG/PGP Key-Id: 0xE704E230 - http://keyserver.linux.it -------- D.Lgs. 196/2003 -------- Si avverte che tutte le informazioni contenute in questo messaggio sono riservate ed a uso esclusivo del destinatario. Nel caso in cui questo messaggio Le fosse pervenuto per errore, La invitiamo ad eliminarlo senza copiarlo, a non inoltrarlo a terzi e ad avvertirci non appena possibile. Grazie. _______________________________________________ Ocfs2-users mailing list Ocfs2-users at oss.oracle.com http://oss.oracle.com/mailman/listinfo/ocfs2-users -------------- next part -------------- An HTML attachment was scrubbed... URL: http://oss.oracle.com/pipermail/ocfs2-users/attachments/20081202/4e275a90/attachment.html
Sunil Mushran
2008-Dec-02 20:11 UTC
[Ocfs2-users] Please urgent help required - OCFS2 and VPN again
The interconnect is your bottleneck. Lorenzo Milesi wrote:> Hi all... > > I already wrote before on the list about the solution I have at a > customer running DRBD8+OCFS2 on two remote sites connected via VPN. > The different suggestions helped improving the situation, but still > we're having big troubles. We've also upgraded the old server with a new > and much more powerful one but there was nearly no improvement at all! > The situation is resumed as: > SITE A: Dual Core 2GHz Pentium, 1Gb ram, 1 SATA hdd for /, 3 SATA hdd in > software raid5, DRBD on /dev/md0. > SITE B: Quad Core 2.4HGz Pentium, 2Gb ram, 3 SATA HDD in software raid5, > DRBD on /dev/md1. > > The two sites are connected using two ADSL, with TWO bonded VPN. > > Both machines run Debian Etch fully updated, kernel 2.6.26-bpo.1-686 SMP > with deadline scheduler, DRBD 8.0.13, OCFS2 1.4.1-1. > The shared data partition is 187G, 30 of which used. > > The recent upgrade to OCFS2 1.4 and kernel 2.6.26 didn't improve the > performances as much as I expected. > > The main problems we have are: > 1. very high load average: this was previously caused by very high > iowait percentages, but with the new server the load is high while top > says the machine is 99-100% idle! > 2. very slow dir browsing: Sunil pointed me to the user guide, where he > talks about inode stat. How can I raise inode cache memory? I've done > several searches without result... The server actually uses less than > 300Mb of ram out of the 1Gb installed... > 3. very long umount time: I often (not always) experience an extremely > long umount time. During the period while the process is executing iftop > says there's a high usage of network transfer. I suppose it's > transfering file locks, but is it possible that stays stuck for more > than one hour, and still going? > > This is the configuration file of OCFS2. The quad-core is file-server-2. > > #/etc/ocfs2/cluster.conf > node: > ip_port = 7777 > ip_address = 192.168.0.1 > number = 0 > name = file-server-1 > cluster = ocfs2 > node: > ip_port = 7777 > ip_address = 192.168.2.31 > number = 1 > name = file-server-2 > cluster = ocfs2 > cluster: > node_count = 2 > name = ocfs2 > > > What is stunning me is that on file-server-2 we run a rsync backup during the night on a local machine on the network, and it takes less than 20m! Doing the same on the other server throws the load average to the stars! > > We're in a critical situation because this solution is deployed since a long time and it's not yet working as expected. > If nobody has suggestion we have no problem in paying qualified support for solving these problems. In this case please contact me directly. > Sunil, can I get Oracle support for this? > > Thank you. >
Lorenzo Milesi
2008-Dec-08 20:13 UTC
[Ocfs2-users] Please urgent help required - OCFS2 and VPN again
Il giorno mar, 02/12/2008 alle 12.11 -0800, Sunil Mushran ha scritto:> The interconnect is your bottleneck.thanks for your answer. what could be the minimum bandwidth for letting ocfs2 work fine? thanks -- Lorenzo Milesi - lorenzo.milesi at yetopen.it YetOpen S.r.l. - http://www.yetopen.it/ C.so E. Filiberto, 74 23900 Lecco - ITALY - Tel 0341 220 205 - Fax 178 607 8199 GPG/PGP Key-Id: 0xE704E230 - http://keyserver.linux.it -------- D.Lgs. 196/2003 -------- Si avverte che tutte le informazioni contenute in questo messaggio sono riservate ed a uso esclusivo del destinatario. Nel caso in cui questo messaggio Le fosse pervenuto per errore, La invitiamo ad eliminarlo senza copiarlo, a non inoltrarlo a terzi e ad avvertirci non appena possibile. Grazie.
Lorenzo Milesi
2008-Dec-09 09:00 UTC
[Ocfs2-users] Please urgent help required - OCFS2 and VPN again
I suppose rsync takes ages in reading all directory tree and directory listing, as it has to dig it all! This shouldn't be affected by drbd "writes", isn't it? Is it due to high latency? thanks! Il giorno lun, 08/12/2008 alle 17.15 -0800, Sunil Mushran ha scritto:> ocfs2 does not have a concept of "primary". While we do use > lower node number as a tie breaker in the dlm, that in itself > should not be affect the performance much. > > See what rsync is doing. > > Lorenzo Milesi wrote: > > Thanks again. > > > > Another curiosity: as I told we have a rsync backup running on one node. > > If that node is set "primary" (node number 0 in cluster.conf) that rsync > > takes up to 4h for transferring very few files, while if I set it > > "secondary" it takes 20m! > > > > > > > > Il giorno lun, 08/12/2008 alle 12.33 -0800, Sunil Mushran ha scritto: > > > >> ocfs2's interconnect usage is latency sensitive. It does not use > >> much bandwidth. While it has small packets, it has lots of them. > >> It has been written with a low latency gige private interconnect > >> in mind. > >> > >> In your setup you are using drbd which is replicating all block > >> writes across the interconnect. That is bandwidth sensitive. > >> > >> The performance of the ocfs2 file system (all fses actually) is partly > >> dependant on the underlying block device. If the block device is > >> slow, the overall performance of the fs will be slow. > >> > >> Lorenzo Milesi wrote: > >> > >>> Il giorno mar, 02/12/2008 alle 12.11 -0800, Sunil Mushran ha scritto: > >>> > >>> > >>>> The interconnect is your bottleneck. > >>>> > >>>> > >>> thanks for your answer. what could be the minimum bandwidth for letting > >>> ocfs2 work fine? > >>> > >>> thanks > >>> > >>> > >>> > >> > >-- Lorenzo Milesi - lorenzo.milesi at yetopen.it YetOpen S.r.l. - http://www.yetopen.it/ C.so E. Filiberto, 74 23900 Lecco - ITALY - Tel 0341 220 205 - Fax 178 607 8199 GPG/PGP Key-Id: 0xE704E230 - http://keyserver.linux.it -------- D.Lgs. 196/2003 -------- Si avverte che tutte le informazioni contenute in questo messaggio sono riservate ed a uso esclusivo del destinatario. Nel caso in cui questo messaggio Le fosse pervenuto per errore, La invitiamo ad eliminarlo senza copiarlo, a non inoltrarlo a terzi e ad avvertirci non appena possibile. Grazie.