原子操作与 STM

Question

I was trying some benchmark over these two ( Software transactional memory & Atomic Operation, i guess both are not same ), though i have not done much using STM ( it seems hard to use ),but i successfully tried counting benchmark ie all the threads are incrementing a shared counter 50mn times, and noticed that atomic operation works better than STM.
So i want to know since STM are not fully developed, do they perform well in realistic situation than Atomic operation?
Has some one switched to other because of performance? please share the information..
related journal i found on the web is Concurrent programming without locks

PS I am using JAVA as programming language. STM:- multi-verse. AtomicOperatinn: AtomicInteger.

Answer 1

Atomic operations and STMs are very different beasts. In particular, atomic operations are much "lower level" constructs - in fact, atomic operations of various kinds are generally used to implement STM. Basically:

• An atomic operation allows you to do a concurrency-safe update of a single object . This is easy and fast, typically just a single instruction on modern hardware.
• STM allows you do to a concurrency-safe update of multiple objects . This is complex.

STM systems therefore need to add additional book-keeping overhead to manage the complexity of a transactional update - so if you only need to update a single object an atomic operation will nearly always be faster.

Also, you should be aware that there is no consensus on the best way to design STM systems - it's an active research topic and there are many tradeoffs involved. So what performs well on one STM system might not perform well on another STM system etc.

For example: my favourite STM system at the moment is Clojure's - it's particularly appealing because it supports multi-version concurrency control and never impedes readers - which gives a significant performance advantage in many common scenarios.