The CMS implementation uses multiple garbage collector threads to scan (mark) the unused objects that can be removed (sweep). This garbage collector is preferable for applications that require short GC pauses, and who can share processor resources with the garbage collector while the application is running.

The CMS algorithm enters into STW mode in only two cases: when objects in Old Generations are still referenced from the thread entry point or static variables, and when the application changed the state of the heap while CMS is running which makes the algorithm go back and reiterate the object tree to validate that it had marked the correct objects.

With this collector, promotion failure is the greatest cause for concern. Promotion failure occurs when a race condition occurs between a collection of objects from the Young and Old Generations. If the collector needs to promote objects from the Young Generation to the Old Generation and there is not enough space, it has to first STW to create the space. In order to make sure this doesn't happen in the case of the CMS collector, increase the size of the Old Generation or allocate more background thread to the collector to compete with the allocation rate.

In order to provide high throughput, CMS uses more CPU to scan and collect objects. It is good for long-running server applications, which are adverse to application freezes. So, if we can allocate more CPU to avoid application pauses, we can choose the CMS collector for GC in our application. To enable the CMS collector, set the -XX:+UseConcMarkSweepGC option.