Daneel Yaitskov
Hi Ben,
Thanks for bothering. I solved issue. That was ancient cabal. I build ghc successfully once I installed Cabal-3. I thought that configure script will warn me about incompatible cabal :)
Hmm, interesting. I'm actually rather surprised that the problem was cabal-install. Afterall, GHC's build system doesn't use cabal-install apart from building Hadrian.
Btw could you shed some light on RTS locking?
I am interested in:
ccall traceUserBinaryMsg(MyCapability() "ptr", msg "ptr", len);
because it looks slower on smaller chunks. Writing data in 64 bytes chunks vs 4k (ghc 8.11) is twice slower.
https://github.com/yaitskov/eventslog-benchmark benchmark [ LoadWithBinary-b1048576-t001-c0064] lasted for cpu 4.23 ms real 4.24 ms benchmark [ LoadWithBinary-b1048576-t001-c0128] lasted for cpu 3.97 ms real 3.91 ms benchmark [ LoadWithBinary-b1048576-t001-c1024] lasted for cpu 2.73 ms real 2.61 ms benchmark [ LoadWithBinary-b1048576-t001-c2048] lasted for cpu 2.34 ms real 2.23 ms benchmark [ LoadWithBinary-b1048576-t001-c4096] lasted for cpu 2.16 ms real 2.20 ms
Eventlog messages tend to be small.
So application developer has to think about bufferization to speed up. OpenTelemetry gains 6x speed improvement with following hack.
... Interesting. Please do open a GitLab ticket to track this. You might also be interested in #17949, which is covers the case of traceEvent being expensive when tracing is disabled.
I see that some EventLog methods use mutexes, but user message function don't. It just appends to buffer and can flush if it is full. So I don't see where locking happening. MyCapability is just arithmetic
#define MyCapability() (BaseReg - OFFSET_Capability_r)
I guess capability locking happens somewhere upper - before C.
There is no locking necessary. Capabilities are owned by a particular operating system thread. The capability is held until the program yields it (which generally only happens during a GC or foreign function call) and during this time the program has full access to all of the state in the Capability structure without the need for a lock. This is why each Capability has its own eventlog buffer; it allows lock-free tracing. Cheers, - Ben