🚀 go-pugleaf

adams@pdv2.fmr.maschinenbau.th-darmstadt.de (Adams) said:

adams> Requirement was, that about 30 frames / second have to be
adams> processed.  Field experience showed that about 2 frames a second
adams> could be retrieved, yielding a transfer rate of 100 kBytes...

adams> The application does nothing else than to read in frames, one
adams> about 50 kBytes, unbuffered (simple read()). One hint: ...  more
adams> than 90% idle, though single user mode!, and drive is kept
adams> selected.

The 90% idle time is not a hint, it is the converse of doing only two
transactions per second (actually you could be having 90% sys time
instead of idle, but I would not believe it!).

Now, let's see. Suppose that the filesystem is so incredibly fragmented
that the system has to do a seek every say 8KBytes; you can do on modern
large fast drives around 50 seeks/second (including rotational latency),
more if seeks are to nearby cylinders: and thus one, even assuming the
drive is constantly seeking, should see at least 400KB/s of bandwidth.

No, something is badly wrong. Are you sure your read(2) calls are in
large units (this matters *a lot*!), say each read(2) is for an entire
50KB frame?

Are you sure you have "properly" setup the filesystem parameters?

Are you sure the disks's and the controller's parameters are also
properly setup (e.g. enabling the read ahead buffer in the disk)?

adams> In article <1992Jul29.221440.1166@ukw.uucp> lupe@ukw.uucp (Lupe
adams> Christoph) writes:

lupe> In this thread, nobody has asked or volunteered if this app is
lupe> using the filesystem or reading from a raw disk. The performance
lupe> sounds like the latter, loosing revs all the time. This would also
lupe> explain the high idle time percentage.

On 5 Aug 92 01:11:47 GMT, adams@pdv2.fmr.maschinenbau.th-darmstadt.de (Adams) said:

adams> You are right, the application stores data in files. It does not
adams> read from raw disk.

But the SGI filesystem is famously efficient in the sense that it
delivers almost all the raw disk bandwidth. Sequential read rates of
well over 1MB/s are easily observed (and 1MB/s at 50KB/frame is already
20 frames/s). Indeed the SGI filesystem is so well done that it delivers
almost all the raw speed in random reads too.

I get well over 300KB/s on a 400 MB SCSI disk reading from the
filesystem on my lowly 386, and I would get more if the SVR4 driver had
less appalling overheads. The original BSD FFS was getting 300KB/s reads
too, or one third of total bandwidth, ten years ago on a VAX-11/780.

That should not be factor!
--
Piercarlo Grandi                   | ARPA: pcg%uk.ac.aber@nsfnet-relay.ac.uk
Dept of CS, UCW Aberystwyth        | UUCP: ...!mcsun!ukc!aber-cs!pcg
Penglais, Aberystwyth SY23 3BZ, UK | INET: pcg@aber.ac.uk