The process of managing memory in the VM is handled by the Allocator and the Garbage Collector (GC). These components operate on an area of memory that is reserved for VM processing called the Java heap. The Allocator assigns areas of the Java heap for Java objects. Objects are considered as live when they have a chain of references to them that start from root references, such as those found in thread stacks. When that reference or pointer no longer exists, the objects are considered as garbage. The role of the Garbage Collector is to manage the storage in the Java heap and reclaim memory by removing garbage. This process is known as Garbage collection.
The Allocator is a component of memory management that is typically overshadowed by the task of garbage collection. However, it is a critical, if not small part, of the overall process. The Allocator manages pools of free memory and how the free memory is consumed. It is also responsible for allocating areas of storage in the Java heap for objects at the request of applications, class libraries, or the VM. Every allocation requires a heap lock to stop concurrent threads trying to access the same area of memory at the same time. When an object is allocated, the heap lock is released. If there is insufficient space to allocate the object, allocation fails, the heap lock is released, and the GC is called. If the GC manages to recover some space on the heap, the Allocator can resume operations. If the GC does not recover enough space, it returns an
Acquiring a heap lock for every allocation would be an intensive operation with a knock on impact to performance. To get around this problem, small objects are allocated to thread local heaps (TLH).
Thread local heaps (TLH)
To improve performance, allocation caches are reserved in the heap for different threads. These allocation caches are known as thread local heaps (TLH) and allow each thread to allocate memory from its cache without acquiring the heap lock. A TLH is typically used for small objects of less than 512 bytes (768 bytes on 64-bit VMs) although larger objects can be allocated from the cache if there is sufficient space.
If a thread allocates a lot of objects, the allocator gives that thread a larger TLH to reduce contention on the heap lock.
A TLH is predefined with an initial default size of 2 KB. On every TLH refresh, the requested size for that thread is increased by an increment (default 4 KB). The requested size can grow up to a predefined maximum (default 128 KB).
After every GC cycle, the TLH requested size for each thread is reduced, sometimes by as much as 50%, to take account of threads that reduce their allocation rate and no longer need large TLHs.
For very inactive threads, the requested size can even drop below the initial value, down to the predefined minimum (512/768 bytes). For very active threads, the maximum TLH requested size might be reached before the next GC occurs.
Larger TLHs can help reduce heap lock contention, but might also reduce heap utilisation and increase heap fragmentation.
The following options control the requested TLH size:
Typically, when the maximum TLH size is increased, you should also increase the increment proportionally, so that active threads can reach the maximum requested TLH size more quickly.
To prevent applications running out of memory, objects in the Java heap that are no longer required must be reclaimed. This process is known as garbage collection (GC). When garbage is collected, the garbage collector must obtain exclusive access to the heap, which causes an application to pause while the clean up is done. This pause is often referred to as a stop-the-world pause because an application must halt until the process completes. In general, the first step in the GC process is to mark the objects that are reachable, which means they are still in use. The next step is to sweep away the unmarked objects to reclaim memory. The last step, which isn’t always required unless the heap has become very fragmented, is to compact the heap.
Eclipse OpenJ9 has a number of GC policies designed around different types of applications and workloads. Picking the right policy very much depends on your usage and performance goals.
Generational Concurrent policy
If you have a transactional application, with many short lived objects, the Generational Concurrent GC policy (
-Xgcpolicy:gencon) is probably best suited, which aims to minimize GC pause times without compromising throughput. This is the default policy employed by the VM, so if you want to use it you don’t need to specify it on the command line when you start your application.
gencon policy, the Java heap is divided into two main areas, the nursery area, where new objects are created and the tenure area, where objects are moved if they have reached tenure age.
The nursery area is subdivided into two further areas, the allocate space and the survivor space. The GC process is illustrated in the following diagram, which shows a sequence of 4 main events:
- Objects are created in the allocate space.
- The allocate space is full.
- A local GC scavenge process runs and reachable objects are either copied into the survivor space or into the tenure area if they have reached tenure age. Any objects that can’t be reached are left untouched and subsequently cleared.
- The allocate and survivor spaces swap roles. The original survivor space becomes the allocate space where new objects are created, and the original allocate space becomes the survivor space ready for the next local GC scavenge process.
The relative sizes of the allocate and survivor spaces are dynamically adjusted by a technique called tilting. When the nursery area is first created, it is evenly divided between the allocate and survivor spaces. If, after a GC scavenge process is run, the amount of space required for the survivor area is comparatively small, the boundary between the two spaces is adjusted by tilting. For example, if the survivor space requires only 10% of the nursery area, the tilt ratio is adjusted to give 90% of the nursery area to the allocate space. With more space available for new objects, garbage collection can be delayed.
The tenure age of an object is determined by the VM and reflects the number of times that an object has been copied between the allocate space and the survivor space. The age is in the range 1 – 14 and is adjusted dynamically by the VM depending on the overall amount of space that is used in the nursery area. For example, if an object has a tenure age of 5, it has been copied backwards and forwards between allocate and survivor spaces 5 times. If the VM sets a tenure age of 5 based on the percentage of space remaining in the nursery area, the next scavenge moves the object from the nursery to the tenure area. You can set an initial tenure age with the
-Xgc:scvTenureAge option. You can also prevent the VM dynamically adjusting the tenure age by setting the
Xgc:scvNoAdaptiveTenure option so that the intial age is maintained throughout the run time of the VM.
Within the tenure area, new objects are allocated into the small object area (SOA), which is illustrated in the earlier diagram (see 3.). A large object area (LOA) is set aside for objects greater than 64 KB that cannot be allocated into the SOA to minimize fragmentation. The LOA is allocated by default but is reduced and removed after a few GC cycles if it isn’t populated. To prevent the creation of an LOA, you can specify the
-Xnoloa option on the command line when you start your application. When the tenure area is close to full a global GC is triggered.
The local GC scavenge reduces pause times by freqently reclaiming memory in the nursery area which, for a transactional application with many short-lived objects, has the most recyclable space. However, over time the tenure area might become full. So, whilst a local GC scavenge process is operating on the nursery area, a concurrent global GC process also runs alongside normal program execution to mark and remove unreachable objects from the tenure area. These two GC approaches combine to provide a good trade-off between shorter pause times and consistent throughput.
A special mode of the
gencon policy is known as Concurrent Scavenge (
-Xgc:concurrentScavenge), which aims to minimize the time spent in stop-the-world pauses by collecting nursery garbage in parallel with running application threads. This mode can be enabled with hardware-based support and software-based support.
Pages: 1 2