Jet runs inside a Java Virtual Machine which uses garbage collected heap memory. Depending on heap size, version of the JVM and the specific garbage collector used garbage collection can suddenly create large pauses in the application, leading to unpredictable spikes in latency and drops in throughput.
To avoid this, Jet supports storing data in
IMap in what's called
native or off-heap memory using the High-Density Memory
feature of Hazelcast. This allows JVM to operate with smaller heaps and
do garbage collection more efficiently, while allowing Jet to store
large amount of data.
Currently off-heap memory support in Jet is limited to data stored in
IMap, the state for running jobs are still in on-heap data structures.
Configuring Off-Heap Memory
You can configure an
IMap to use off-heap memory by configuring
hazelcast.yaml. First, you need to dedicate a specific portion of the
physical memory available to off-heap and then enable off-heap memory
for the specific maps. The node needs to have enough free physical
memory for this feature to work. For example, to dedicate
memory for off-heap and to enable it for the
IMap with name
huge-map, you can make the following configuration changes in
hazelcast: native-memory: enabled: true size: unit: GIGABYTES value: 128 map: huge-map: in-memory-format: NATIVE
For additional configuration options, see configuring high-density memory section of the Hazelcast manual.
Off-heap memory can also make use of persistent memory technologies such as Intel Optane. Depending on the value size, this offers almost similar performance to memory at a lower cost, making it attractive for storing large amounts of data efficiently.
Please refer to the configuring persistent memory section of the Hazelcast manual for details on how to enable persistent memory.