ITEM: CS7711L

LIBRARY: Sysback restore performance on SP switch.

Problem: When using sysback over the SP high-speed switch,
 performance is extremely poor in one or both directions.

Solution:
* Verify that *all* traffic goes through the high speed switch and not
  just one direction.  Do this by verifying that the hostnames used for
  the client and the server in the sysback configuration on both machines
  are the hostnames associated with the switch interface.
* Check the error report for errors on the interface
  `errpt -a` to look for css0 switch errors
* Test for network congestion by sending a largish file both directions
  between the affected nodes.
* Check system load average with `w` or check idle with iostat.  Idle
  under 30% or so, or load above 4 could cause problems.
* Test disk and tape performance locally to each machine to verify that
  this is not due to bus latency or over-loaded disk (dd, backup, etc).
* For tape drives, use a fixed block size (262144 for magstar, 1024 for
  8mm or 512 for QIC or DAT) rather than variable (zero).
* Check the Maximum Transmission Unit size:
  `netstat -in` MTU should be 65520
  `no -a | grep tcp_mssdflt`  should match MTU
* check these `no` values (Network Options):
  `no -a | more`
    `thewall` should usually be around 16384 or 32768
    `tcp_send_space` should be the same as or larger than the largest
     packet to be sent (ie, same as the sysback buffer size)
    `tcp_recv_space` should be 2x `tcp_send_space`
    `RFC1323` should be 1 to handle packets over 64k
* Check these interface values on each machine:
  `lsattr -El css0`
    `rpool_size` should be the same as, or a whole multiple of the `no`
     value `tcp_recv_space`
    `spool_size` should be double `rpool_size`
  These values should be the same for ALL machines involved.  When
  combined, they cannot equal more than 16m because of pinned memory
  limitations
* The following steps will need to be taken to make changes to an active
  switch adapter:
    `cd /usr/lpp/ssp/css`       Go to the right directory.
    `./ifconfig css0 down`      Bring down the interface.
    `./ifconfig css0 detatch`   Fence the node.
    Wait about 20 seconds to allow the other switch ports notice the
     detachment of this node.
    `chgcss  -l css0 -a rpoolsize=\#\#\#\#\#\# -a spoolsize=\#\#\#\#\#\#\#`
    `./rc.switch`               Start the interface back up.
* Additional Notes:
  * To check switch adapter stats, here are the netstat equivalents:
    `/usr/lpp/ssp/css/estats`
    `/usr/lpp/ssp/css/vdiddle3`
  * General network performance tips:
    If tcp window is smaller than mtu, then runs synchronously.
      (ie. one in, one out, no buffering.
    If a packet is under 4k, it's network buffer comes from `thewall`
    Larger packets come from rpool and spool
    If mtu is smaller than packet size, there must be enough memory
      allocated in the rpool to store *all* fragments of the packet,
      plus any other inbound network traffic, until the fragements can
      be reassembled for processing.

Response:

Closing with Customer Approval



Support Line: LIBRARY: Sysback restore performance on SP switch. ITEM: CS7711L

Dated: September 1998 Category: N/A

This HTML file was generated 99/06/24~13:30:17

Comments or suggestions?
Contact us