Once the critical-heap-percentage is exceeded, operations do not complete, and instead throw a LowMemoryException for Region.put(), Region.create(), Region.putAll(), any index creation operation, and at the execution of any function. This continues until the tenured heap usage falls more than 2% below the configured critical-heap-percentage, at which point the system returns to normal operation without intervention.

Exceeding this threshold implies that the member is about to run out of tenured heap memory, and the Java garbage collector has not been able to reduce the quantity of allocated memory. Since further heap allocations would likely cause an OOME, the member does not complete those operations that require further memory allocations. Exceeding this threshold signals that a significantly larger than expected burst of allocations occurred or that something is critically wrong with the member. If not a burst, then a non-trivial fix or redesign will be needed. Something is not right in the design or in the sizing of the jvm for the current set of regions that this member hosts.

Be aware that exceeding the threshold on one member may affect other members. With operations not completing, those other members may believe that the member exceeding the threshold has crashed; they may then take further action and revoke the membership of the member that is above the critical-heap-percentage. This can cause the remaining members to increase their usage of memory to compensate for the missing member. If these remaining members are not sized to handle the increased load, the critical heap threshold issue may cascade through the system.

Symptoms May Include

  • Operations that would require memory allocation throw LowMemoryException, and what happens after that depends on the code that catches the exception.

  • With system monitoring, alerts identify a member that is above the specified critical-heap-percentage.

Action to Take

Temporary Measures

  • Force garbage collection. From the command line, find the PID of the member, and use it in the command
    jcmd <pid> GC.run

  • Proactively shut down the member that is above the critical-heap-percentage, and restart it with a larger tenured heap size. This action does an orderly shutdown, assuming that an orderly shutdown is preferred over a crash due to an OOME. It also presumes that the member can be given more heap space (-Xmx and -Xms). This measure does not address the underlying question of why the member surpassed the configured critical-heap-percentage; that still needs to be done.

  • If the member that is above critical-heap-percentage is hosting considerably more buckets of partitioned regions than other members hosting the same partitioned regions, a rebalance may help. It moves entire buckets from one member to another, thereby reducing the heap space load. Invoke the rebalance on all partitioned regions, once connected, with gfsh rebalance, or use the --include-region option to explicitly specify those regions that are to be included in the rebalance.

Redesign Actions

Redesign falls into 3 categories. More than one action in more than one category may be appropriate to mitigate the issue.

  1. Increase the amount of available memory. The increase may need to be applied to all the members.

    • Start the member with a larger tenured heap size.  Upon restart of the member, increase values for -Xmx and -Xms.

    • Increase the critical-heap-percentage. This implies that there really is more heap space available, and it is currently wasted.

    • Use off-heap memory.

    • Increase the quantity of members (servers) hosting partitioned regions to horizontally scale the system.

  2. Tune garbage collection.

    • Lower the initiating occupancy fraction of the CMS algorithm, such that the concurrent collection cycle is initiated earlier and at a lower threshold of heap space usage. This will hopefully free tenured heap space such that the system will not reach the critical-heap-percentage.

  3. Modify the system's regions so that less heap space is used.

    • Evict data. Set an eviction threshold in the resource manager, and change the region types to those with an lru-heap-percentage eviction strategy. Decide what happens when an entry is evicted. It can be destroyed (entire entry) or the value can be overflowed to a disk.

    • Expire data.

  • No labels