topic Re: ext2 optimizations for writes? in Operating System - Linux

ext2 optimizations for writes?

Jim Turner — Wed, 29 Dec 2004 11:22:02 GMT

I have a DL380G3 running RHEL 3.0 ES. He's on a private GigE LAN and SAN-attached to a couple of MSA1500 disk arrays. His sole purpose in life is to accept Data Protector backup data over GigE and write it to files (virtual tapes) on the MSA1500's.

I'm currently using ext2 filesystems on the MSA1500's since I'm interested in flat-out write performance (vs. journaling or read performance). What recommendations can you kind folks make for tuning ext2 for max write efficiency or for replacing ext2 with something else that's HP-supportable under RedHat Enterprise Linux 3.0 ES?

Thanks!
Jim

Re: ext2 optimizations for writes?

Steven E. Protter — Wed, 29 Dec 2004 11:28:26 GMT

ext2 and ext3 are both journaled filesystems designed to mimic the vxfs/jfs functionality we are used to on hpux.

I don't either of them are very well optimized for writes but they are well supported. Without outright replacing the filesystem the best bet is to try and mimic async write functionality like we have on hpux.

Thats probably done with mount options.

This link kinda gets into the topic:
http://www.beowulf.org/archive/2002-April/006630.html

Maybe something else in my little google hunt will help.

http://www.google.com/search?hl=en&lr=&q=ext2+ext3+%22mount+options%22+%22write+performance%22&btnG=Search

SEP

Re: ext2 optimizations for writes?

Jim Turner — Wed, 29 Dec 2004 11:32:18 GMT

Steven,

Are you sure ext2 is a journaling file system? I thought that was the major diff between ext2 and ext3.

Thanks,
Jim

Re: ext2 optimizations for writes?

xyko_1 — Wed, 29 Dec 2004 13:16:20 GMT

Jim,

as far as I know, ext3 is ext2 improuved by journaling feature. So ext2 is not "journal protected", then it is supposed to have more i/o performance.

I'm affraid that doesn't have good chances to improuve performance by tunning ext2.

regards,
xyko

Re: ext2 optimizations for writes?

Steven E. Protter — Wed, 29 Dec 2004 16:26:08 GMT

I think the last post is correct and mine is incorrect. ext3 added journaling.

My links should provide you the mount options you need to maximize write reliability or throughput.

SEP

Re: ext2 optimizations for writes?

dirk dierickx — Thu, 30 Dec 2004 01:53:09 GMT

ext2 is _not_ journalised, ext3 was build upon ext2 and has journaling.

A no brainer mount option is 'noatime', which means the access time on files will not be updated.

Re: ext2 optimizations for writes?

Vitaly Karasik_1 — Thu, 30 Dec 2004 02:40:03 GMT

Jim,

I suggest you to start from global overview of your system:

-hardware [MSA500 firmware/disks speed/RAID configuration]
- linux kernel & storage drive tuning
- choosing optimal fs parameters (block size, ...)

Re: ext2 optimizations for writes?

Jim Turner — Thu, 30 Dec 2004 10:09:06 GMT

Thank you all for the responses so far. In reverse order ...

Vitaly, that's what I started doing yesterday. I have large diagram on my dry-erase board right now that shows data flow (and cmd/ctrl/status) between HP-UX client, HP-UX DP 5.1 cell mgr, and Linux device server (talks to the MSA1500's). I can run 500GB of client data to a single LTO-2 (Ultrium 460) in 2 hrs 53 mins. That same client data takes 6 hrs 55 mins to write to the MSA1500 disks. Obviously, I want disk backup to run faster than tape backup which it is not at this point. I've got to figure out the right combination of the following:
+ MSA1500 array and virtual drive parameters
+ Linux filesystem/mount parameters
+ Linux Fibre Channel HBA settings
+ Linux TCP/IP settings
+ Data Protector "virtual" tape drive settings

Dirk, thanks for the ext2/ext3 confirmation. I have remounted the MSA filesystems with noatime, but from what I've read it looks like that only benefits read operations. It didn't seem to yield a measurable improvement to sequential writes.

Steven, thanks for the follow-up on ext2/ext3.

Xyco, as with the others, thanks for the ext2/ext3 journaling confirmation.

When I built the ext2 filesystems, I took all the defaults except for largefiles (one inode per 1MB) and sparse superblock (didn't need three zillion copies of the superblock since the eight filesystems are 1.2TB each). It's likely I've missed an opportunity for optimization at the filesystem level. These are simple 1.2TB mountpoints with about 291 each 4GB files that Data Protector treats as virtual tapes.

I'm also shooting in the dark with the Data Protector virtual tape drive settings. Currently they are:
+ Block size (kB) = 256
+ Segment size (MB) = 800
+ Disk agent buffers = 16
These are the settings that give me blazing performance on Ultrium 460 tape drives, but they may be grossly inappropriate for "virtual" tape drives writing to virtual tapes (disk).

This really has become quite a challenge trying to take a box full of general purpose parts (Linux server, MSA1500's, SATA shelves) and make them work faster than a monolithic device (Ultrium 460) designed specifically for backup purposes.

I welcome everyone's continued suggestions.

Thanks,
Jim