Performance Luxun vs. Apache Kafka

#Performance Benchmark - Luxun vs. Apache Kafka

Note, in the following test, when flush is enabled, following setting are used by Luxun and Kafka brokers.

log.flush.count=10000
log.default.flush.interval.ms=1000
log.default.flush.scheduler.interval.ms=1000

Apache Kafka(0.7.2) Tests on single machine

In the following tests, one producer process, one consumer process and one broker process are all run on my server(CentOS 6.3, HD with EXT4), and there is only one active topic on broker for testing in each round. These tests are performed to check the pure performance of Apache Kafka messaging system by separating the impact of network. See full hardware spec for benchmark here.

JVM setting:

producer : -Xmx1g

consuemr : -Xmx1g
broker : -Xmx4g

###Test Suite 1 - Standalone Async Producing(batch size = 200, flush enabled)

Test Case	messge size(bytes)	thread number	nMsgs sent	throughput(MB/s)	speed(nMsg/s)
TC11	100	1	2000000	8.9108	93436.1131
TC12	1024	1	2000000	34.6360	35467.2814
TC13	2048	1	2000000	41.8233	21413.5055
TC14	4096	1	2000000	45.4029	11622.9064
TC15	100	10	2000000	10.8317	113578.2838
TC16	1024	10	2000000	50.8123	52031.8435
TC17	2048	10	2000000	66.2953	33943.1791
TC18	4096	10	2000000	77.9504	19955.3001

###Test Suite 2 - Standalone Async Producing(batch size = 200, flush disabled)

Test Case	messge size(bytes)	thread number	nMsgs sent	throughput(MB/s)	speed(nMsg/s)
TC21	100	1	2000000	30.7985	322945.2608
TC22	1024	1	2000000	47.1473	48278.8587
TC23	2048	1	2000000	45.0782	23080.0300
TC24	4096	1	2000000	45.3011	11597.0938
TC25	100	10	2000000	58.8870	617474.5292
TC26	1024	10	2000000	75.1520	76955.6351
TC27	2048	10	2000000	71.7283	36724.8756
TC28	4096	10	2000000	78.9102	20200.9999

###Test Suite 3 - Standalone Sync Producing(flush enabled)

Test Case	messge size(bytes)	thread number	nMsgs sent	throughput(MB/s)	speed(nMsg/s)
TC31	100	1	2000000	9.7101	101817.4413
TC32	1024	1	2000000	35.6540	36509.6751
TC33	2048	1	2000000	41.2870	21138.9676
TC34	4096	1	2000000	43.1964	11058.2771
TC35	100	10	2000000	11.0854	116238.5214
TC36	1024	10	2000000	49.6070	50797.5211
TC37	2048	10	2000000	68.1505	34893.0528
TC38	4096	10	2000000	77.3768	19808.4523

###Test Suite 4 - Standalone Sync Producing(flush disabled)

Test Case	messge size(bytes)	thread number	nMsgs sent	throughput(MB/s)	speed(nMsg/s)
TC41	100	1	2000000	46.5207	487804.8780
TC42	1024	1	2000000	48.6797	49847.9637
TC43	2048	1	2000000	46.2936	23702.2991
TC44	4096	1	2000000	44.8036	11469.7314
TC45	100	10	2000000	81.1983	851426.1388
TC46	1024	10	2000000	81.8234	83787.1806
TC47	2048	10	2000000	65.9851	33784.3545
TC48	4096	10	2000000	69.0455	17675.6518

###Test Suite 5 - Standalone Consuming(fetch size = 1024 * 1024bytes)

Test Case	messge size(bytes)	thread number	nMsgs fetched	throughput(MB/s)	speed(nMsg/s)
TC51	100	1	2000000	91.7436	962000.9620
TC52	1024	1	2000000	95.4607	97751.7107
TC53	2048	1	2000000	97.6465	49995.0005
TC54	4096	1	2000000	100.4616	25718.1802

Luxun 0.6.0 Tests on single machine

In the following tests, one producer process, one consumer process and one broker process are all run on my server(CentOS 6.3, HD with EXT4), and there is only one active topic on broker for testing in each round. These tests are performed to check the pure performance of Luxun messaging system by separating the impact of network. See full hardware spec for benchmark here.