libera/#clasp - IRC Chatlog
Search
13:45:37
drmeister
That's the estimated time to complete processing 48GB of sequencing data on hermes (my linux box) and on bigmac (a 36 core iMacPro).
13:46:39
drmeister
I think this reflects "false sharing" that occurs when lots of threads access the queue lock or the lock I placed on the hash table that accumulates results.
13:48:32
drmeister
What's frustrating about this is profiling and flame charts don't show this problem.
14:02:42
drmeister
I've been frustrated by the perf tool on linux. Maybe I need to put more time into it.
14:48:27
drmeister
It's partly because seqan has this function and perf mem report prints the entire freaking function name.
14:50:33
drmeister
Ugh - I can't be bothered to find it again. But I'm looking at a report like this:
14:51:14
drmeister
Let's build a tool for performance measurements that can't handle C++ function names.
14:54:22
Bike
oh, weird, so the names aren't actually long, it just puts a ton of spaces in for no reason
14:54:39
Bike
or i guess there are other lengthy names elsewhere and it wants to assure table columns align
15:17:12
Bike
hm, so, it looks like on linux, if i link libunwind in there's a crash (std::terminate) at what is probably the first throw... even if we don't actually make any calls into libunwind
17:13:17
drmeister
Some of the names are very long - like 1500 characters. I wrote a script to fold everything into a more reasonable width.
17:17:36
Bike
like if you had another variable bound after sequences, if that value signaled an error/whatever sequences wouldn't get cleaned up because the unwind-protect wouldn't be entered