Hazelcast with Azul Zing Benchmarks

Benchmark

Table of Contents

Key Findings
- 1GB Heap Results Summary
- 2GB Heap Results Summary
Hotspot
- Hotspot – 2 Node Cluster, 1GB Heap, 49% Heap Fill
- Hotspot – 2 Node Cluster, 2GB Heap, 49% Heap Fill
Zing
- Zing – 2 Node Cluster, 1GB Heap, 49% Heap Fill
- Zing – 2 Node Cluster, 2GB Heap, 49% Heap Fill

Overall RadarGun configuration

Test Duration – 15s Warmup, 20m Stress Test
Entry Sizes – 15%: 1024KB, 35%: 2048KB, 50%: 4096KB
Get / Put Ratio – 67% / 33%
2 node cluster

Key Findings

Zing reduces full GC pauses.

So we would expect this to reflect in average latencies and, because larger full GCs are avoided,
much more so in the largest max latencies. And this is what we found.

The average was reduced by 40%. The 99.99th percentile max latency was only 8ms for Zing
but 55ms for HotSpot.

We conclude that Hazelcast® with Azul Zing JVM has lower latency and much lower variability
for more predictable latencies. This in turn allows Hazelcast to fit into tighter response time SLAs.

These results are applicable to small heap sizes where Hazelcast is storing small amounts of data of 500MB to 1.5Gb
per node and are relevant to Hazelcast and Hazelcast Enterprise.

Note that in Hazelcast Enterprise HD, we store data off heap and only run with a small heap.
So a similar benefit to one in these tests with 1 to 2GB heaps would be realized.

See below for summaries of results for 1GB and 2GB heap cases and the extensive test results further down.