The hardware transaction is implemented in cache. A potential problem is that variables involved in a single transaction cannot fit into cache. The solution is to let processors abort such a transaction with an abort handler. However, for transaction with the number of variables that is essentially too large for cache, no matter how many times it tries, it always gets aborted. The transaction then is livelocked. A way to avoid this livelock is to use short transactions with acceptable number of variables in memory.
This comment was marked helpful 0 times.
pwei
Aborting of a transactional memory commit will be somewhat costly due to the time it will take to reload the cache line and restore the values in the transaction. This is not an earth-shattering amount of time, but can build up if it happens too often. One can also imagine that the more this memory commit gets stuck, the more it holds up the rest of the program. If too many processors are trying to commit the same places in memory (which hopefully doesn't happen or is prevented by other hardware), then a particular processor may never be able to commit its changes.
The hardware transaction is implemented in cache. A potential problem is that variables involved in a single transaction cannot fit into cache. The solution is to let processors abort such a transaction with an abort handler. However, for transaction with the number of variables that is essentially too large for cache, no matter how many times it tries, it always gets aborted. The transaction then is livelocked. A way to avoid this livelock is to use short transactions with acceptable number of variables in memory.
This comment was marked helpful 0 times.
Aborting of a transactional memory commit will be somewhat costly due to the time it will take to reload the cache line and restore the values in the transaction. This is not an earth-shattering amount of time, but can build up if it happens too often. One can also imagine that the more this memory commit gets stuck, the more it holds up the rest of the program. If too many processors are trying to commit the same places in memory (which hopefully doesn't happen or is prevented by other hardware), then a particular processor may never be able to commit its changes.
This comment was marked helpful 0 times.