Performance

The Krati distribution has included a number of tests for collecting performance statistics. If you have a Krati distribution with versions 0.3.4 and above, you can simply run the command below to collect read/write throughput and latency numbers on your own computer.
ant test.loggc -Dtests.to.run=TestDataCache

If you want to evaluate DataCache with MappedSegment and ChannelSegment, you can run the following commands respectively.
ant test.loggc -Dtests.to.run=TestDataCacheMapped
ant test.loggc -Dtests.to.run=TestDataCacheChannel

The Krati distribution also includes other peroformance tests on data store classes. You might want to play with these tests to get a general idea of Krati performance.

Test Configuration

This page provides a quick glance of Krati performance. The performance figures were collected using the setup below:

• Krati Configuration
• 1 Writer
• 4 Readers
• Data Size: 0.5~2 KB, Avg. 1 KB
• Batch Size: 10,000
• Segment Size: 256 MB
• Member Count: 5,000,000 (typical partition size)
• Test Machine
• Mac OS X Version: 10.5.8
• Processor: 2 x 2.26 GHz Quad-Core Intel Xeon
• Memory: 24 GB 1066 MHz DDR3
• Startup Disk: Macintosh HD
• Java 6
• Sun Hotspot JVM
• -server -Xmx16G

Read/Write Throughput

The write throughput is approximately 20~30 writes per millisecond for MemorySegment and roughly 10~20 writes per millisecond for MappedSegment and ChannelSegment. The persistency and recovery achieved via disk files and redo logs have an impact on write throughput.

The read throughput is approximately 1000~1200 writes per reader thread per millisecond for MemorySegment. It is an order of magnitude faster than throughput obtained using MappedSegment or ChannelSegment.

Read/Write Latency

The ChannelSegment has the highest write latency. As shown in the figure below, approximately 80% of writes have a latency between 10 and 50 microseconds. The write latencies for MemorySegment and MappedSegment are approximately on the same level with the majority of writes finished between 1 and 10 microseconds.

The read latency for MemorySegment is under 1 microsecond. In contrast, over 95% of reads from ChannelSegment range between 10 and 50 microseconds. MappedSegment is in the middle.

For an in-depth comparison of Krati and BDB JE, please refer to A Thorough Look at Krati vs. BDB JE - A Comparison of Throughput, Latency and GC.