HP XC System Software Release Notes for Version 2.1

12.3.4 ELAN TRAP Queue Error Seen on Some Quadrics MPI Applications

Some QsNet

MPI applications that generate many concurrent DMA operations

might encounter the following error:

ELAN TRAP -0- Unknown - Queue Error

This error terminates the program, which is believed to be caused by high rates

of ELAN PutGet operations.

It is possible to work around this problem by setting the LIBELAN_PUTGET_THROT-

TLE environment variable to a value lower than its default value of 32.

12.3.5 The qsnet Database May Contain Entries to Nonexistent Switch

Modules

Depending on your system topology, the qsnet diagnostics database may contain

entries to nonexistent switches.

This issue is manifested as errors reported by the /usr/bin/qsctrl utility

similar to the following:

# qsctrl

qsctrl: failed to initialise module QR0N03: no such module (-7)

...

In the previous example, the switch_modules table in the qsnet database is

populated with QR0N03 even though the QR0N03 module is not physically present.

This problem has been reported to Quadrics, Ltd.

To work around this problem, delete the QR0N03 entry (and any other nonexistent

switch entries) from the switch_modules table, and restart the swmlogger

service:

# mysql -u root -p qsnet

mysql> delete from switch_modules where name="QR0N03";

mysql>quit

# service swm restart

In addition to the previous problem, the IP address of a switch module may be

incorrectly populated in the switch_modules table, and you might see the

following message:

# qsctrl

qsctrl: failed to parse module name 172.20.66.2

...

Resolve this issue by deleting the IP address from the switch_modules table

and restarting the swmlogger service:

# mysql -u root -p qsnet

mysql> delete from switch_modules where name="172.20.66.2";

mysql>quit

# service swm restart

_________________________ Note _________________________

You must repeat the previous procedure if you rerun the

cluster_config utility because the qsnet database is recreated

during a cluster_config operation.

Interconnect Notes 12-3