IPU-POD128 build and test guide
IPU-Link traffic test
The traffic test acts as a smoke test for all IPU-Links of a cluster before deploying applications.
The traffic test can be invoked with the
--traffic
option. Note that for a traffic test to pass, a prerequisite is that
the IPU-Link and IPU-Link training tests have passed.
$
vipu-admin
test
cluster cluster1 -traffic
+---------+----------+--------+---------------------------------------------------+
| Test
| Duration | Passed |
Summary
|
+=========+==========+========+===================================================+
| Traffic | 92.23s
| 3/4
| Traffic test failed
|
+---------+----------+--------+---------------------------------------------------+
|
|
|
| Errors encountered in traffic test 1
|
+---------+----------+--------+---------------------------------------------------+
|
|
|
| corrected link errors: 460
|
+---------+----------+--------+---------------------------------------------------+
|
|
|
| error counter IPU-Link 1 in ipum1, IPU '1' is 250 |
+---------+----------+--------+---------------------------------------------------+
|
|
|
| error counter IPU-Link 1 in ipum4, IPU '1' is 210 |
+---------+----------+--------+---------------------------------------------------+
test cluster (cluster1): failed: Some tests failed.
This example shows a situation where the IPU-link traffic test has failed due to too many correctable errors being
detected. Should this occur please try reseating the IPU-Link cables associated with the referenced IPU-M2000
units. If that does not resolve the issue, please contact Graphcore support.
Version: latest (2021-11-25)
84