Heap: Sizing and Swapping

A common problem is configuring a heap that is too large. You have a 64 GB machine—​and by golly, you want to give Elasticsearch all 64 GB of memory. More is better!

Heap is definitely important to Elasticsearch. It is used by many in-memory data structures to provide fast operation. But with that said, there is another major user of memory that is off heap: Lucene.

Lucene is designed to leverage the underlying OS for caching in-memory data structures. Lucene segments are stored in individual files. Because segments are immutable, these files never change. This makes them very cache friendly, and the underlying OS will happily keep hot segments resident in memory for faster access.

Lucene’s performance relies on this interaction with the OS. But if you give all available memory to Elasticsearch’s heap, there won’t be any left over for Lucene. This can seriously impact the performance of full-text search.

The standard recommendation is to give 50% of the available memory to Elasticsearch heap, while leaving the other 50% free. It won’t go unused; Lucene will happily gobble up whatever is left over.

There is another reason to not allocate enormous heaps to Elasticsearch. As it turns out, the JVM uses a trick to compress object pointers when heaps are 30.5 GB or less.

In Java, all objects are allocated on the heap and referenced by a pointer. Ordinary object pointers (OOP) point at these objects, and are traditionally the size of the CPU’s native word: either 32 bits or 64 bits, depending on the processor. The pointer references the exact byte location of the value.

For 32-bit systems, this means the maximum heap size is 4 GB. For 64-bit systems, the heap size can get much larger, but the overhead of 64-bit pointers means there is more wasted space simply because the pointer is larger. And worse than wasted space, the larger pointers eat up more bandwidth when moving values between main memory and various caches (LLC, L1, and so forth).

Java uses a trick called compressed oops to get around this problem. Instead of pointing at exact byte locations in memory, the pointers reference object offsets. This means a 32-bit pointer can reference four billion objects, rather than four billion bytes. Ultimately, this means the heap can grow to around 32 GB of physical size while still using a 32-bit pointer.

Once you cross that magical 30.5 GB boundary, the pointers switch back to ordinary object pointers. The size of each pointer grows, more CPU-memory bandwidth is used, and you effectively lose memory. In fact, it takes until around 40–50 GB of allocated heap before you have the same effective memory of a 30.5 GB heap using compressed oops.

The moral of the story is this: even when you have memory to spare, try to avoid crossing the 30.5 GB heap boundary. It wastes memory, reduces CPU performance, and makes the GC struggle with large heaps.

It should be obvious, but it bears spelling out clearly: swapping main memory to disk will crush server performance. Think about it: an in-memory operation is one that needs to execute quickly.

If memory swaps to disk, a 100-microsecond operation becomes one that take 10 milliseconds. Now repeat that increase in latency for all other 10us operations. It isn’t difficult to see why swapping is terrible for performance.

The best thing to do is disable swap completely on your system. This can be done temporarily:

sudo swapoff -a

To disable it permanently, you’ll likely need to edit your /etc/fstab. Consult the documentation for your OS.

If disabling swap completely is not an option, you can try to lower swappiness. This value controls how aggressively the OS tries to swap memory. This prevents swapping under normal circumstances, but still allows the OS to swap under emergency memory situations.

For most Linux systems, this is configured using the sysctl value:

vm.swappiness = 1 

Finally, if neither approach is possible, you should enable mlockall. file. This allows the JVM to lock its memory and prevent it from being swapped by the OS. In your elasticsearch.yml, set this:

bootstrap.mlockall: true