Hacker News new | ask | show | jobs
by gpderetta 3658 days ago
Transactions compose and TM is potentially easier to use.

Also some algorithms necessarily require locking multiple mutexes (think for example of a wait-for-any operation). HTM in principle would allow implementing them more efficiently.
1 comments
pizlonator 3658 days ago
Transactions don't compose with I/O or any OS side-effects that the TM system isn't in the loop on, while locks compose fine with those things. One module that uses locks is free to call another module that uses locks so long as the relationship is one-way like user->kernel, so locks compose also, just with different constraints. In short, transactions don't compose with I/O while locks don't compose with cyclic relationships. So, the statement that "transactions compose" is false in both directions: it's false because transactions don't compose in general, and it's false because it implies that locks don't, while they in fact do.

The only empirical evidence about transactions being easier to use is actually contrary: even with the availability of HTM, it's not used nearly as widely as locks, because if you enable HTM for every lock then everything runs a lot slower (or doesn't run at all).

link
gpderetta 3658 days ago
Under the proposed C++ TM, synchronised blocks can do i/o. They are not transactional, but do compose.

If you have two locked data structures and you want an operation to manipulate both atomically you need to expose the locks. And if you want the operation itself to compose with others... well get hairy.

TM is widely used I believe by least Haskel and Closure programmers, even with non hardware accelerated versions because of the ease of use.

HTM is another story, it's availability is still relatively restricted. Time will tell.

link
pizlonator 3657 days ago
Closure is indeed TM-based, but that's not an A/B test of ease-of-use. I don't think there exists evidence to suggest that transactions are easier to use.

I think you are confused about the meaning of composability. The synchronized/atomic TM proposal for C++ is roughly single-global-lock. This means that you can still deadlock if you use synchronized to protect IPC with another process, which then does an IPC call back to you, which you service on another thread that also uses synchronized. Whenever these transactional proposals fall back on locking, they immediately bring in the pitfalls of locking.

The C++ TM proposal is a joke. There's no good reason to use concurrency if you won't get a speed-up, and the current design of synchronized is sure to make it more difficult to get speed-ups.

link