HTM or fine-grained locking

[danghvu]

In the original paper: https://www.cs.princeton.edu/~mfreed/docs/cuckoo-eurosys14.pdf

It mentions there are two modes: HTM or fine-grained locking. If I'm using the default configuration, which is actually being used ? I run it on a Intel machine (x86_64, a Xeon server), does that mean it enables HTM by default ?

[manugoyal]

Hey @danghvu! The paper does indeed explore both HTM and fine-grained locking, but for this implementation, we decided to stick with just fine-grained locking. The implementation is intended to work well for all kinds of C++ types across a variety of machines, so we settled on fine-grained locking, which allows safe memory management for all C++ types.

It would definitely be possible to support HTM for intel machines for specific key-value pairs (integer-integer, for example), but we haven't gotten around to that yet.

[ldalessa]

I was wondering if there's any easy way to test HLE on the current fine-grained locking version. Is there any obvious way to get the spinlock to play along? I presume it would have something to do with the std::atomic_flag test-and-set.

[manugoyal]

It might be possible. Found this blog post on intel's website, where someone implements an elided mutex with HLE: https://software.intel.com/en-us/blogs/2013/07/25/fun-with-intel-transactional-synchronization-extensions.

So presumably, yes, spinlock can be modified to use HLE and still support the lock and unlock functions. Haven't tried it out myself, though.