ROOT I/O Meeting

Friday 5 May 2017, 16:00 → 17:25 Europe/Zurich

VIDYO

Attendees: Peter, Philippe, Enrico, Guilherme, Pere, Vassil.

Vassil, Peter, Enrico: ras.

David: Could PR#523 be merged in for 6.10

Philippe: yes, of course.

David: high luminosity pill ups, 8000 events, 600GB total, high I/O.  o far I have not seen any significant improvement I need to run more test. Replay with software then hardware.  1000s of buffers.  pick different sizes.   I am working on getting good setup and launch it soon.

Philippe:  While investigating the crashing during GDML parsing, I concluded that this was due to a kernel bug where memory pages that moved back from the page file *sometimes* get the executable bit (or some other page table information) copied incorrectly and/or not on time.  We have verified that the problem is seen for both bare-metal and VM running a 3.10 kernel and is *not* seen in neither bare-metal and VM running a 4.11.  Some of the changes in the kernel source code between the two version are explicitly race condition in the page fault/retiring handler that lead to already moved page being written to.

Guilherme: Included suggestion of returning nice error when client is kept beyond the lifetime of the server.  Cannot use exception
Factory returns a shared_ptr and the server keeps a weak_ptr and when the server goes down it checks that all client are gone.  We have no spinning threads (using condition variable).   Server created and deleted the server-thread.  Client are obtain through factory method.  Still want to improve/reduce the memcpy from client to server.    Helper functions/classes are defined in unnamed namespace.  The only things left is the termination condition (nullptr added to the queue by the server destructor).  

Guilherme: Not sure where to put the flush.

Guilherme: I still need to write ‘standalone’ test.

Guilherme/Vassil/Philippe: technical discussion on the merge request.

16:00 → 16:40 Round Table