From: Chris Dos Subject: Re: Millions of files and directory caching. Date: Wed, 23 Oct 2002 16:26:17 -0600 Sender: nfs-admin@lists.sourceforge.net Message-ID: <3DB72209.5060901@chrisdos.com> References: Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii; format=flowed Return-path: Received: from 12-252-254-16.client.attbi.com ([12.252.254.16] helo=windwalker.chrisdos.com) by usw-sf-list1.sourceforge.net with esmtp (Exim 3.31-VA-mm2 #1 (Debian)) id 184TxN-0000ny-00 for ; Wed, 23 Oct 2002 15:26:29 -0700 Received: from localhost (localhost.localdomain [127.0.0.1]) by windwalker.chrisdos.com (Postfix) with ESMTP id 4E56F19C401 for ; Wed, 23 Oct 2002 16:26:25 -0600 (MDT) Received: from chrisdos.com (unknown [192.168.10.127]) by windwalker.chrisdos.com (Postfix) with ESMTP id 0688A19C0AF for ; Wed, 23 Oct 2002 16:26:17 -0600 (MDT) To: nfs@lists.sourceforge.net Errors-To: nfs-admin@lists.sourceforge.net List-Help: List-Post: List-Subscribe: , List-Id: Discussion of NFS under Linux development, interoperability, and testing. List-Unsubscribe: , List-Archive: The RAID 10 makes much better sense that the RAID 0+1 that I've seen before. I never realized that there was such an incredible performance hit using RAID 5. I knew there was some, but not to that degree. RAID 5 would be horrible for a mail server or web server. But RAID 5 would probably be good for files that are larger than the stripe size. I thought about getting a NVRAM drive and putting the journal it. But this box is at full capacity and I don't have any other drive bays free for the drive. But right now, performance has increased so much, that having the journal RAID 10 isn't imapacting me at this time. Are there any other patches besides the htree patch do you think I should apply to increase my performance. Will moving to NFS over TCP give me much of an improvement? I am using Gigabit on the server and only 100BT Full Duplex on the clients. Chris Craig I. Hagan wrote: >>I've been able to get 320 threads so far with the 2.4.19 kernel. I just >>didn't know what the limit was. I have a feeling, that by the time this >>is done, I'll need to set it up for 512 threads. >> >>I'd love to be able to make a bunch of smaller RAID1's. But the size of >>my mount points just won't allow it. But when you create a 3ware RAID >>10, it does seem to make a bunch of small RAID 1's. Not the typical >>RAID 10 setup that I've seen. Though, that might be the way it displays >>it's RAID 10 config. > > > well, there are two ways of doing raid with mirrors and stripes. > > 10: > > one stripe across several mirror pairs. This is technically > the best as it gives you the most amount of fault tolerance. > nearly all hardware raid solutions do this. Many software > raid solutions will let you do this. > > picture: > > stripe is veritical, mirrors hozontal > > stripe0,d1: [mirror1, d1 + mirror1, d2] > stripe0,d2: [mirror2, d1 + mirror2, d2] > stripe0,d3: [mirror3, d1 + mirror3, d2] > stripe0,d4: [mirror4, d1 + mirror4, d2] > > so the striping driver sees four logical disks, each of which are a mirror > pair. the advantage of this is that you can (if lucky) sustain a maximum of > four single disk faults (one per mirror) and still have the volume operational > in a degraded state. > > more imporantly, a single disk fault will degrade only the performance of one > of the stripe pieces (1/4 of the total space will have reduced (50%) read > capacity). Write performance to that stripe should be of similar performance. > I'm omitting reconstruction costs. Odds state that it is unlikely that your > next disk fault will be the other disk in that same mirror set (for the case of > double faults). > > I've also seen folks do it as raid 0+1 (veritas on solaris often > does this, making many folks feel that this is 'right'): > > you make two stripes and mirror them, e.g. > > mirror is now veritical, stripes horizontal > > mirror1, disk1: stripe1,d1 + stripe1,d2 + stripe1, d3 + stripe1, d4 > mirror1, disk2: stripe2,d1 + stripe2,d2 + stripe2, d3 + stripe2, d4 > > Here, a single disk fault will immediately kill 50% of your total read > performance. You now have a high probability chance that your next fault will > take out your other stripe. > > Additionally, reconstruction time is bound by how long it takes to resilver the > entire volume rather than 1/4 (for the example) of it. > > >>As for my setup, the results are finally in. To sum it up, RAID 5 is >>bad, RAID 10 is good. Performance went up by almost one full order of > > > if you read up on what raid5 has to do for the case of a small file write, it > would make sense. the small write is turned in to a "read, modify, recrc, > write" across a full raid5 stripe. > > so, if you use a chunk size of 32k and have 8+1 raid, any write smaller than > 256k (32k*8 data disks) requires that all 256k be read in, the relevant bits > changed, a new crc computed, and everything thrown back rather than just > crc'ing the data and writing. This requirement to operate on much larger chunks > of data is what makes raid5 poor for an NFS server. > > >>Now if I can just come up with a way to make it do snapshots like a >>Netapp I'd be set. I heard that the Linux Volume Manager has this >>capability. Does anyone know if this is true? Back in 1999 there as a >>snapfs project that was started, but I don't think it ever got off the >>ground. > > > i've used the snapshot ability of LVM. be wary of how you mount the snapshots, > i think you need to mount them as ext2 as it will try to do some journal ops > and will get *very*pissed* that it can't write to the block device. > > also, you should seriously consider two things: > > * the htree patch. this makes it so that you can put 10^7th files > in an ext3 directory and have it perform very well > > * a NVRAM block device (and put your journal on this). This will > allow the nfs server to more readily handle write requests. > failing the nvram block device, you may > want to journal to a dedicated disk. > > >> Chris >> >>pwitting@Cyveillance.com wrote: >> >>>One thing I found with RH 7.3 is that it would not start more than 128 NFS >>>threads, even if I told it to start 256, I would get no more than 128. You >>>Can confirm this by running >>> >>> ps aux | grep -c [n]fs >>> >>>Of course, your custom kernel may behave differently, my customizations was >>>a simple patch for jfs and a recompile to scan all LUNS (FC connected IBM >>>ESS/Shark). >>> >>>And as for RAID issues; its very common for folks to automatically go with >>>RAID 5 because of its efficiency, and its speed during sequential reads. >>>However, there's lots of occasions where it performance can be awful, >>>particularly when performing random writes. RAID 10 is very good when one >>>needs fast continuous space, but there are downsides to RAID 10 this as >>>well. The big problem with both of these solutions is that they act as a >>>single "spindle", when two request arrive in parallel, they must be handled >>>serially, with the heads jumping back and forth between data1 and data2. If >>>the data can be laid out across 4 RAID 1 arrays, Array1's heads stay reading >>>data1 and Array2's head stay reading data2. The net result in more speed. >>>DB@, and I assume Oracles and other high end databases actually make >>>assumptions about "containers" being on different "spindles" and will lay >>>the data out to maximize throughput. (Before moving to the Shark, my >>>database had 66 RAID 1 pairs) This results in even less data efficiency, >>>since it can be hard to evenly spread the data out, and may take a bit more >>>work to maintain (then again, it can be easy, depending on the application), >>>But in a server environment cranking lots of transactions, it can be a big >>>performance win. >>> >>>From: Chris Dos >>> >>>>I wasn't able to get the tranfer done last night. Hopefully it'll >>>>finish by tonight and I can let everyone know the results tomorrow. >>>> >>>>Does anyone know what is the maximum number of NFS threads I can run. >>>>It seems like 256 wasn't enough for my platform, so I've upped it to >>>>320. What is the absolute limit I can go? >>>> >>>> Chris >>>> >>>>Lever, Charles wrote: >>>> >>>> >>>>>hi chris- >>>>> >>>>>someone recently reported to me that oil companies like RAID 10 >>>>>because RAID 5 performance is terrible for NFS servers. seems >>>>>like you are on the right path. >>>>> >>>>> >>>>> >>>>> >>>>> >>>>>>Man, you setup seems extreamly close to mine and it's even >>>>>>for the same >>>>>>type of business. Let me give you a run down on what I have, >>>>>>and what >>>>>>I've been doing. >>>>>> >>>>>>We had a EMC Symetrix SAN/NAS that held 5.7 TB worth of disk. >>>>>>We were >>>>>>only using about 550 GB of it, so the decision was to move >>>>>>away from the >>>>>> EMC because of ongoing support cost and complexity issues. >>>>>>The EMC >>>>>>was working in a NAS configuration serving files via NFS, and it was >>>>>>also sharing some of it's disk to a Sun 420R which was then serving >>>>>>files via NFS. >>>>>> >>>>>>The clients are a mix of Solaris 2.6/7.0/8.0 and Redhat 7.2 with the >>>>>>stock kernel. There are three RedHat 7.2 servers (two are updated >>>>>>running the 2.4.18-17 kernel, the other run is running the >>>>>>2.4.7 kernel) >>>>>>that serve mail, Two Solaris servers that serve web, and one Oracle >>>>>>server that did have external disk to the EMC. The clients are >>>>>>connected to the switch at 100BT FD. The clients use the following >>>>>>mounting options: >>>>>>udp,bg,intr,hard,rsize=8192,wsize=8192 >>>>>> >>>>>>The server built to replace the EMC is built as follows: >>>>>>Hardware: Tyan 2462 motherboard with five 64 bit slots >>>>>> Dual AMD 2100+ MP Processors >>>>>> 2 GB PC 2100 RAM >>>>>> Two 3Ware 64 bit 7850 8 port RAID cards >>>>>> 16 Maxtor 160 GB Hard Drives >>>>>> 3 Netgear 64 bit 621 Gigabit cards >>>>>>Software: >>>>>> Redhat Linux 7.3 (All patches applied) >>>>>> Custom 2.4.19 Kernel (I also rolled one using the >>>>>> nfs-all patch and the tcp patch, but Oracle didn't >>>>>> like starting it's database when mounted to it. >>>>>> Don't know why) >>>>>>Config and Stats: >>>>>>This server was configured using each 3ware RAID card in a >>>>>>RAID 5, and >>>>>>then mirrored via software RAID to the other 3Ware RAID card. >>>>>>This gave >>>>>>us 1.2 TB of usable space. I've moved the data off the EMC to this >>>>>>server, and I had to move the Oracle servers Oracle database to this >>>>>>server as well and export via NFS. (The Oracle (Sun 420R) server can >>>>>>only hold two drives) The server is connected to the network via >>>>>>Gigabit FD. >>>>>>I'm running 256 NFS threads, and even then, that doesn't seem like >>>>>>enough according to the output of /proc/net/rpc/nfsd. What's >>>>>>the limit >>>>>>on the maximum number of threads I can run? >>>>>> >>>>>>Ouput of /proc/net/rpc/nfsd: >>>>>>rc 21716 821664 85468545 >>>>>>fh 233099 88275136 0 233232 6591060 >>>>>>io 2359266294 1934417182 >>>>>>th 256 19536 1012.250 787.570 875.870 1350.400 840.930 220.190 96.470 >>>>>>50.970 20.000 229.690 >>>>>>ra 512 4558260 16517 7803 5421 4228 3293 2686 1808 1279 1270 923900 >>>>>>net 86311942 86311942 0 0 >>>>>>rpc 86311925 17 16 1 0 >>>>>>proc2 18 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 >>>>>>proc3 22 100 47898973 4984 2138510 29020056 120 5542814 390404 117607 >>>>>>130 0 9 163535 8 103856 62847 516154 205343 17118 772 60 128525 >>>>>> >>>>>>All the exports from the server have these options: >>>>>>rw,no_root_squash,sync >>>>>>And I'm passing these options to the kernel: >>>>>>/bin/echo 2097152 > /proc/sys/net/core/rmem_default >>>>>>/bin/echo 2097152 > /proc/sys/net/core/rmem_max >>>>>>/bin/echo "65536" > /proc/sys/fs/file-max >>>>>> >>>>>>And even with all of this, I'm having issues with this box. >>>>>>The client >>>>>>NFS mounts are extreamly slow. So slow, that some services >>>>>>time out and >>>>>>the daemon stops all together. This is a very bad thing. So I've >>>>>>pulled my hair out, beat my head against the wall, and contemplated >>>>>>using a sledge hammer to finish this slow painful death. I've tried >>>>>>connecting the server via 100 FD instead of Gig to see if >>>>>>that would fix >>>>>>the problem, nada. So, I got around to thinking, that when I had >>>>>>configured older 3Ware 6500 RAID cards in a RAID 5 on another server, >>>>>>performance sucked. Converting it to RAID 10 solved that issue. The >>>>>>RAID 5 performance was supposed to be fixed in the 7500 series,but I >>>>>>suspected it was not. So I decided to explore this tangent >>>>>>and pull a >>>>>>couple of all nighters to make this happen. So.... >>>>>> >>>>>>I broke the Software RAID, reconfigured one of the >>>>>>controllers as RAID >>>>>>10, giving me 640 GB of space. Started copying as much data >>>>>>as I could >>>>>>between 10pm-7am Saturday and Sunday night, and as of this >>>>>>morning, I've >>>>>>been able to move 1/4 of the mail (one full mount) to the RAID 10. >>>>>>Already performance of my NFS has increased. Customers aren't >>>>>>complaining now about slow mail (or no mail access at all for that >>>>>>matter). After tonight I should have all the mail moved over to the >>>>>>RAID 10 and I should be able to give you an update tomorrow. If >>>>>>everything goes as planned, I'll move the web sites the next day, and >>>>>>then this weekend, I'll reconfigure the controller that is in >>>>>>a RAID 5 >>>>>>config, to a RAID 10, and then bring up the software RAID 1 >>>>>>between the >>>>>>controllers. >>>>>> >>>>>>So, I think your problem is caused by RAID 5 and not NFS, >>>>>>just like mine >>>>>>is. I'll know more tomorrow. >>>>>> >>>>>>If anyone can see anything wrong with my configs, or other >>>>>>optimizations I can make, please let me know. This is a very high >>>>>>profile production environment. I need to get this thing running >>>>>>without a hitch. >>>>>> >>>>>> Chris Dos >>>>>> >>>>>>Matt Heaton wrote: >>>>>> >>>>>> >>>>>>>I run a hosting service that hosts about 700,000 websites. >>>>>>> We have 2 >>>>>>>NFS servers running Redhat 7.2 (2.4.18 custom kernal, no >>>>>>>nfs patches). The servers are 850 GIGS each (IDE RAID 5). >>>>>>>THe clients >>>>>>>are all 7.2 Redhat with custom 2.4.18 kernels on them. My >>>>>>>question is >>>>>>>this. I believe lookups/attribs on the files and directories are >>>>>>>slowing down performance considerably because we literally have 4-5 >>>>>>>million files on each nfs server that we export. One of >>>>>>>the NFS servers >>>>>>>is running EXT3 and the other is XFS. Both work ok, but >>>>>>>under heavy >>>>>>>loads the clients die because the server can't export stuff fast >>>>>>>enough. The total bandwidth out of each NFS server is LESS than 10 >>>>>>>Mbit. The trouble is that I am serving a bunch of SMALL >>>>>>>files. Either >>>>>>>I am running out of seek time on my boxes (IDE Raid 850 GIGS per >>>>>>>server), or it is taking forever to find the files. >>>>>>> >>>>>>>Here are my questions. >>>>>>> >>>>>>>1) Can I increase the cache on the client side to hold the entire >>>>>>>directory structure of both NFS servers? >>>>>>> >>>>>>>2) How can I tell if I am just maxing the seek time out on >>>>>>>my NFS server? >>>>>>> >>>>>>>3) Each NFS server serves about 60-100 files per second. >>>>>>>Is this too >>>>>>>many per second? Could I possibly be maxing >>>>>>>out seek time on the NFS servers? My IDE Raid card is the >>>>>>>3ware 750 >>>>>>>with 8 individual IDE ports on it. >>>>>>> >>>>>>>4) Is there anything like cachefs being developed for >>>>>>>linux?? Any other >>>>>>>suggestions for persistent client caching for NFS? >>>>>>>Free or commercial is fine. >>>>>>> >>>>>>>Thanks for your answers to some or all of my questions. >>>>>>> >>>>>>>Matt >>>>>> >>> >>> >>>------------------------------------------------------- >>>This sf.net email is sponsored by: Influence the future >>>of Java(TM) technology. Join the Java Community >>>Process(SM) (JCP(SM)) program now. >>>http://ads.sourceforge.net/cgi-bin/redirect.pl?sunm0002en >>> >>>_______________________________________________ >>>NFS maillist - NFS@lists.sourceforge.net >>>https://lists.sourceforge.net/lists/listinfo/nfs >>> >> >> >> >>------------------------------------------------------- >>This sf.net email is sponsored by: Influence the future >>of Java(TM) technology. Join the Java Community >>Process(SM) (JCP(SM)) program now. >>http://ads.sourceforge.net/cgi-bin/redirect.pl?sunm0002en >> >>_______________________________________________ >>NFS maillist - NFS@lists.sourceforge.net >>https://lists.sourceforge.net/lists/listinfo/nfs >> > > ------------------------------------------------------- This sf.net email is sponsored by: Influence the future of Java(TM) technology. Join the Java Community Process(SM) (JCP(SM)) program now. http://ads.sourceforge.net/cgi-bin/redirect.pl?sunm0002en _______________________________________________ NFS maillist - NFS@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/nfs