Re: 2.3.x wish list? (Block sizes)

Paul M. Hirsch (pauldoom@webcreate.net)
Tue, 25 May 1999 11:39:55 -0500 (CDT)


On 23 May 1999, Stefan Monnier wrote:

> >>>>> "Paul" == Paul M Hirsch <pauldoom@webcreate.net> writes:
> > Solaris and Irix (XFS) offer block sizes from 4KB to 8KB, and from 512B
>
> My MH mailbox would look like:
> 70MB with 1KB-blocks
> 77MB with 2KB-blocks
> 96MB with 4KB-blocks
>
> Since MH mailboxes are about as bad a case as it gets (apart possibly
> from big symlink-farms), it seems that 2KB is a safe default while 4KB
> incur a non-negligible space cost.

Oh, so we are worried about space now? ;) I should have said
"4KB blocks can offer significantly better performance, with a (sometimes
huge) space cost."

> I don't know whether Irix's XFS has fragments (I believe it doesn't), but it
> does have "file in the inode" (not just for symlinks but also for files and
> directories) which might also reduce the need for small blocks.

Ah, data in with the metadata. At my last job they were writing a high
perf fs for supercomputers that used this extensivly. Metadata
was set up in small blocks, data in huge blocks (with fragment support),
then each could be put on seperate RAID volumes, triped, etc. There were
seperate metadata and data caches, and it switched automaticly to direct
disk IO for large enough file sizes. Between the metadata caching and
small metadata block sizes, small IO screamed. And with the big data
block sizes and direct IO, big data screamed as well. Of cource, this fs
will probably never see the light of day. They were working on a Solaris
port when I left. (If all this sounds confusing and/or stupid, excuse my
ignorance. I am no fs hacker. I'm just relating what interesting
facts I can retrieve from my defective memory.)

Basicly, it was an fs that tied together alot of good high perf ideas from
other fs's. I wouldn't want to run it on my laptop, though.

-Paul

------------------------------------------------------------
Paul M. Hirsch Network Engineer EXi, Corp.
smtp-> phirsch@exicorp.com http-> www.exicorp.com
------------------------------------------------------------

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/