Subject: Re: softdep?
To: Brian C. Grayson <bgrayson@orac.ece.utexas.edu>
From: Don Lewis <Don.Lewis@tsc.tdk.com>
List: current-users
Date: 03/29/1999 01:44:53
On Mar 26, 11:11pm, "Brian C. Grayson" wrote:
} Subject: Re: softdep?

}   Here's my pipe-dream reason for the question:
}   I've wondered what performance would be like if one made a
} homebrew NVRAM (sitting on a parallel port or ISA, rather than
} PCI or a SIMM slot) and changed all FFS mounts to be async --
} how much state needs to be kept in NVRAM to be able to recover
} in the case of a panic or power loss?  I don't know much about
} file systems, but only metadata needs to be saved, right?  And
} that shouldn't be _too_ much, especially if it gets cleaned out
} of the NVRAM once the async write of the metadata is done....

I suspect that parallel port would be too slow and even ISA would
hurt performance.  Many (most?) disk writes would need to update
NVRAM.  Ideally the NVRAM should be close to main memory speed.

If you just want to speed up a local filesystem, then just caching
the metadata should be enough, and I don't think it would require
too much NVRAM.  If you want to accelerate NFS writes like
PrestoServe does, then more NVRAM would be required, since the
point is to preserve data written by the NFS clients across a
server crash and the server can't send its responses to the client
until the client's data has been written to stable storage (disk
or NVRAM).

As far as local filesystem performance is concerned, soft updates
has been slightly faster than async mounts in the testing that
I have done.