Memory leak when using the Ganglia plugin in artdaq
The last couple of days I've found that the lion's share of lbne-artdaq memory leaks aren't due to the NetMonInput leak, but rather appear when ganglia is used - i.e., when the FHiCL snippet
is used. For the record, ganglia v3_7_1 is the one in use. I've attached plots of the anonymous page usage for 35ton's runs 18689 and 18690 on both lbnedaq1 (running the RCE fragment generators) and lbnedaq6 (running the eventbuilders and aggregators); the only difference between them is inclusion (run 18689) and exclusion (run 18690) of the above snippet.
#1 Updated by Eric Flumerfelt over 4 years ago
- Status changed from Assigned to Closed
- % Done changed from 100 to 90
- Estimated time set to 2.00 h
I found that the problem was in the code I originally copied out of the gmetric executable. I ran
valgrind --leak-check=full --show-reachable=yes gmetric -n "Test_Metric" -v 0 -t int8 -u "Tests"
to find the location of the memory leaks. After that, I created an initialization routine which isolated the creation of the Ganglia config object and the udp send channels so that it only happens once. This seems to have closed the leak in my testing.