[Info-vax] Sudden problems with slow sftp transfers and slow disk accesses

[Jan-Erik Soderholm]

gwilliams at cfa.harvard.edu wrote 2014-02-18 05:42:
> Cluster of 8 Alpha boxes running V8.3 + patches, recently moved
> from one building to another.  Since the move, we've been experiencing
> odd behaviors: very slow network access (via sftp) and slow disk IO.
>
> Disk storage is mostly on three eternal disk boxes (five three-member
> shadow sets).  No errors are reported via SHOW DEV DSA.
>
> Network cards on all machines are set to 100 MB, full duplex,
> non-autonegotiate, connected via an 8-port GB switch.
>
> An sftp from one of our machine to a local Linux system transferred 288 KB
> of a ~ 6MB file in the first second, the current rate is now down to 5KB/s.
> MONITOR PROCESS/TOPCPU doesn't show the process getting even 1% of the
> CPU and there is nothing else running on the system.  Another attempt on
> the same file transferred 1.5 MB in the first second, then dropped to
> < 50 KB/s.
>
> SHOW MEMORY doesn't show any problems.
>
> A filing operation merging two large files took a matter of seconds
> when both files were on a locally-attached disk.  When both files were
> on a shadow set, the merge took 6+ minutes.  MONITOR LOCK while
> running the latter test showed ENQ/DEQ rates < 1.
>
> Image activation is slow.  It can take several seconds to begin
> running an .exe stored on the shadow set.
>
> MCR SCACP SHOW LAN_DEV/ALL showed numerous errors occurring over the
> past 24 hours, so this evening we replaced the switch connecting these
> 8 machines.  Errors are continuing to appear.
>
> What am I missing?
>
> Gareth
>

Make sure that the switches that your Alphas are connected to
are also set to 100 MB, full duplex, *non-autonegotiate* !
If set to "autonegotiate" they will probably default back to
half-duplex and that will give you huge network problems.

(I had the network techie fix this on our prod server just a
few days ago. Everything else works OK, but am FTP *to* this
server runs realy slow with a lot of collisions on the line.)

Replacing the switches doesn help, of course, as long as they
still runs with "autonegotiate"...

Jan-Erik.