manual:known_issues
no way to compare when less than two revisions
Differences
This shows you the differences between two versions of the page.
— | manual:known_issues [2016/09/01 07:13] (current) – created zenke | ||
---|---|---|---|
Line 1: | Line 1: | ||
+ | ====== Known Issues ====== | ||
+ | |||
+ | * Spike-loss and random crashes experienced with with process numbers that were not a power of two. (OpenMPI 1.4.3) | ||
+ | * Random communication freezes for large messages (i.e. caused by synchronization in the network -> many spikes in short time intervals). | ||
+ | * Still getting random freezes when running v0.8.0-beta on Ubuntu 14.04 LTS with the stock openmpi (Open MPI 1.6.5) | ||
+ | |||
+ | I traced the freezes to SyncBuffer at high firing rates or high synchrony. In most cases it seems as if some processes do not recover from MPI_Allgather when resending the buffer after an overflow. To reproduce run | ||
+ | |||
+ | < | ||
+ | mpirun -n 4 ./ | ||
+ | </ | ||
+ | |||
+ | or for instance | ||
+ | |||
+ | < | ||
+ | mpirun -n 4 ./ | ||
+ | </ | ||
+ | |||
+ | Note, that -n 2 seems to work and 3 or 5 cause other weird problems. Looks like a bug in openmpi. | ||
+ | When compiled against MPICH these problems do not occur. |
manual/known_issues.txt · Last modified: 2016/09/01 07:13 by zenke