STMX presented at 7th European Lisp Symposium

Posted on Updated on

STMX was presented at 7th European Lisp Symposium (ELS 2014) in a technical paper titled High performance concurrency in Common Lisp – hybrid transactional memory with STMX.

Slides, audio recording and video of STMX presentation are available from ELS 2014 website.

Thanks everybody who joined this great event in Paris!

Advertisements

Hardware transactions supported

Posted on Updated on

Since version 1.9.0, STMX supports hardware memory transactions in addition to classic software ones. It uses Transactional Synchronization Extensions (TSX) available at least on the following Intel x86_64 processors released in June 2013:

  • Intel Core i5 4570
  • Intel Core i5 4670
  • Intel Core i7 4770

To actually use hardware memory transactions with STMX, you will need:

  • a CPU supporting Intel TSX, for example one from the above list
  • a 64-bit OS (currently tested on Debian GNU/Linux x86_64)
  • a 64-bit installation of Steel Bank Common Lisp (SBCL) version 1.0.55 or later Note: x86_64 is often named AMD64 – they are the same thing
  • the latest STMX version – download it from GitHub as described in Installation and loading below

The current hardware memory transactions implementation is very fast, yet it still has room for optimizations. Currently, it can accelerate short transactions up to 4-5 times while seamlessly falling back on software transactions when the hardware limits are exceeded. Experiments with hand-optimized code (not yet included in STMX) show that the maximum possible performance increase is 7-8 times.

New Library included: SB-TRANSACTION

Posted on Updated on

Since version 1.3.3, STMX also includes SB-TRANSACTION, a standalone library that does not depend on STMX and provides hardware-only memory transactions on CPUs that support Intel TSX instructions. It is a low-level library providing raw access to the new CPU instructions for hardware transactions. Its performance reaches the theoretical peak supported by the underlying CPU, and it is obviously faster than STMX – it is usually even faster than hand-optimized compare-and-swap fine grained locking code (see benchmark results in doc/benchmark.md). The reason is that it avoids the overhead and the software compatibility requirements of STMX, providing only the raw features – and the limitations – of the underlying CPU. At the moment, SB-TRANSACTION only works on SBCL running in native 64-bit mode on a CPU with hardware transaction support (see the list above).