linuxdcpp-team team mailing list archive

[maksis <1909861@xxxxxxxxxxxxxxxxxx>]

Looks like the benchmarks that you linked use 4-16 KB block sizes for
reading, which is far from optimal when doing regular sequential file
reads for larger files.

"Do you have any idea why in your Ubuntu 20.04 setup, 4MB buffers
performed notably worse than 1MB buffers?"

I'm not really sure, but hardware on that system is quite low-end
compared to the others. Celeron N3150 has a 2 MB L2 cache while the Core
i5-1038NG7 on MacBook Pro has a 6 MB cache (and Ryzen 2700X has even
larger caches). Can that make a significant difference when calculating
TTHs (or with the file reads)?

-- 
You received this bug notification because you are a member of
Dcplusplus-team, which is subscribed to DC++.
https://bugs.launchpad.net/bugs/1909861

Title:
  FileReader is not thread safe on Linux

Status in DC++:
  Fix Committed

Bug description:
  FileReader::readMapped currently modifies the global SIGBUS handler in
  order to catch read errors:

  https://sourceforge.net/p/dcplusplus/code/ci/default/tree/dcpp/FileReader.cpp#l289

  Since the function can be called concurrently from different threads
  (currently hashing/queue recheck/sfv check in DC++) and each of them
  sets and resets the SIGBUS handler, there's a high risk that the
  application will crash in case of read errors as they aren't being
  handler properly.

  More information about the caveats:
  https://www.sublimetext.com/blog/articles/use-mmap-with-care

  These issues are much more likely to happen with AirDC++ as it uses
  multiple threads for hashing. Read errors caused rather nasty crashes
  with corrupted stack traces for one user, but luckily he was able to
  catch the SIGBUS signal with gdb.

  I didn't even spend time in trying to figure out how to make the
  mapped reads work properly, as based on my testing the basic
  FileReader::readCached function is noticeably faster:

  readMapped: 671 files (21.70 GiB) in 9 directories have been hashed in 4 minutes 21 seconds (84.87 MiB/s)
  readCached: 671 files (21.70 GiB) in 9 directories have been hashed in 3 minutes 58 seconds (93.08 MiB/s)

  FileReader::readMapped is now disabled in AirDC++, as I can't see any
  benefits from using it. The included setjmp.h header is even causing
  issues when using clang for compiling on Linux:
  https://bugs.gentoo.org/731676

To manage notifications about this bug go to:
https://bugs.launchpad.net/dcplusplus/+bug/1909861/+subscriptions