35 lines
1.4 KiB
Markdown
35 lines
1.4 KiB
Markdown
|
# Distributed traces (MPI)
|
||
|
|
|
||
|
|
The ovni trace is designed to support concurrent programs running in different
|
||
|
|
nodes in a cluster. It is often the case that the monotonic clock
|
||
|
|
(`CLOCK_MONOTONIC`) are not synchronized between machines (in general they
|
||
|
|
measure the time since boot).
|
||
|
|
|
||
|
|
To generate a coherent Paraver trace, the offsets of the clocks need to be
|
||
|
|
provided to the emulator too. To do so, run the `ovnisync` program using MPI on
|
||
|
|
the same nodes your workload will use. If you are using SLURM, you may want to
|
||
|
|
use something like:
|
||
|
|
|
||
|
|
% srun ./application
|
||
|
|
% srun ovnisync
|
||
|
|
|
||
|
|
!!! warning
|
||
|
|
|
||
|
|
Beware that you cannot launch two MPI programs inside the same srun session,
|
||
|
|
you must invoke srun twice.
|
||
|
|
|
||
|
|
By default, it will generate the `ovni/clock-offsets.txt` file, with the
|
||
|
|
relative offsets to the rank 0 of MPI. The emulator will automatically pick the
|
||
|
|
offsets when processing the trace. Use the ovnisync `-o` option to select a
|
||
|
|
different output path (see the `-c` option in ovniemu to load the file).
|
||
|
|
|
||
|
|
Here is an example table with three nodes, all units are in nanoseconds. The
|
||
|
|
standard deviation is less than 1 us:
|
||
|
|
|
||
|
|
```
|
||
|
|
rank hostname offset_median offset_mean offset_std
|
||
|
|
0 xeon01 0 0.000000 0.000000
|
||
|
|
1 xeon04 1165382584 1165382582.900000 135.286341
|
||
|
|
2 xeon05 3118113507 3118113599.070000 180.571610
|
||
|
|
```
|