Skip to content

Reduce memory consumption

Roel Aaij requested to merge memory_consumption into master

The main cause of large host memory consumption (around something like 3 GB + 1.5 GB / thread) is the memory allocated for secondary vertex monitoring. The initial allocation has been reduced from the theoretical maximum number of secondary vertices to maximum 500 per event.

This is larger than the largest number of secondary vertices in 5000 Bs->Phi Phi events and therefore still conservative. If a batch needs more memory the memory allocated in the HostBuffers will automatically be resized.

A secondary cause of memory usage was the miscalculation of required memory size in the BinaryProvider and MDFProvider, which has been improved. Some latent bugs that showed up along the way have also been fixed.

Closes #158 (closed)

Merge request reports