diff options
| author | Matthias Vogelgesang <matthias.vogelgesang@kit.edu> | 2015-07-07 17:01:37 +0200 |
|---|---|---|
| committer | Matthias Vogelgesang <matthias.vogelgesang@kit.edu> | 2015-07-07 17:01:37 +0200 |
| commit | 9a99cd39053ba9646e618663bdf3cc72c7890fa7 (patch) | |
| tree | b72df267be9562d7f309ea3c2b2309c9cd1024d5 /docs | |
| parent | ed71be36a5694ed2583e4bdf2061aad10b648e1a (diff) | |
Document NVIDIA multi-GPU performance issue
Diffstat (limited to 'docs')
| -rw-r--r-- | docs/manual/using/cluster.rst | 11 |
1 files changed, 11 insertions, 0 deletions
diff --git a/docs/manual/using/cluster.rst b/docs/manual/using/cluster.rst index 7496b25..e266f2c 100644 --- a/docs/manual/using/cluster.rst +++ b/docs/manual/using/cluster.rst @@ -39,3 +39,14 @@ mode, you have to prepare the scheduler:: sched = Ufo.Scheduler(remotes=remotes) sched.set_remote_mode(Ufo.RemoteMode.REPLICATE) sched.run(graph) + + +Improving small kernel launches +=============================== + +UFO uses a single OpenCL context to manage multiple GPUs in a transparent way. +For applications and plugins that require many small kernel launches, multi-GPU +performance suffers on NVIDIA systems due to bad scaling of the kernel launch +time. In order to improve performance on machines with multiple GPUs it is +strongly advised to run multiple ``ufod`` services with differently chosen GPUs +and ports. |