Linux Today is not responsible for the content of the message below.
C Tierney - Subject: Poorly written article, good point, clueless replies posted ( May 12, 2008, 20:21:28 )
To the first few replies to this article, have you ever had to build a multi-GB/s filesystem that can handle abitrary workloads and stay up at least 99.9% of the time? Henry has. I have complaints about his article, but he brings up good points: 1) Other filesystems besides ext3 and XFS aren't supported or tested as well necessary for the types of loads he is describing (and ZFS/Fuse is alpha code). XFS is very good, but the biggest Linux vendor (Redhat) doesn't even support it. 2) Yes, you can fix Linux code yourself, but filesystem are hard. You can't expect some random code jocky to pickup the kernel source and undertsand filesystems. What about the XFS+NFS bug that existed in Linux around 2004 ( http://www.linux.sgi.com/archives/xfs/2004-06/msg0 0100.html) It took over a year for SGI to fix the problem. Open source worked, because the original patch came from a guy at Sony in Japan, but for a year there was silent corruption on any XFS filesystem that was exported via NFS. He says that an LT04 tape drive can push 240 MB/s. Now put 20 of those drives in your system (4.8 GB/s). Now, design your filesystem so that you have extra capacity so that you can interact with the filesystem while the tape drives are banging away (4.8x2=9.6GB/s). This is much more than your home software-raid setup to store your mp3s and pictures. This is High-performance I/O. This is a very common-steup at many of the HPC centers around the world (except they may not be using LTO drives, but enterprise drives). Expecting Linux to push data at that rate is a stretch. However, it is getting better. I would still rather use Linux than bring in one Solaris server that I have to hire or retrain staff because we migrated to Linux a decade ago.
