🚀 go-pugleaf

In article <l7deddINNt72@boogie.cs.utexas.edu>, wilson@cs.utexas.edu (Paul Wilson) writes:
>
>I'm getting some unexpectedly good performance numbers with some
>software I'm running, which looks like it's because file system
>block reads are faster than virtual memory pageins.
>
>Can anybody tell me offhand what's going on?  I'm reading 4K
>blocks from files, and it's unexpectedly fast.  My guess is
>that the filesystem is actually fetching 8K blocks, where the
>VM system is fetching 4K pages, or maybe the file system is
>doing one-block lookahead prefetching.
>
>We're using ULTRIX 4.1.
>
Some of this depends on the hardware configuration.  If you're
doing the sequential reads, then the file system will start doing
read-ahead once it believes you are indeed doing sequential I/O.
So, you read a 4 KB block, it starts that one, starts another for
the next 4 KB of the file and then waits for the first one.  The
2nd will end up the buffer cache and when you read it, no I/O
is required.

If you're using any of most modern SCSI disks they are probably
doing at least one track's worth of read into an internal buffer
and transfering from there instead of reading from the disk.  This
is where most the high read performance comes from.  ULTRIX V4.3
should have file system performance enhancements similiar to the
ones Sun added recently.

If you are doing sequential reads of fairly large files, then
you may want to reconsider a file system with an 8 KB block size
and appropriately larger I/O size.
>
>   Paul
--
Alan Rollow				alan@nabeth.cxo.dec.com