[jira] [Commented] (KAFKA-4971) Why is there no difference between kafka benchmark tests on SSD and HDD?
[ https://issues.apache.org/jira/browse/KAFKA-4971?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15956205#comment-15956205 ] Dasol Kim commented on KAFKA-4971: -- It might be a silly question, If kafka uses the page cache of the OS, if the OS gets bottleneck, the OS is installed in the SSD on the server having both the SSD and the HDD, and the kafka is installed on each of the SSD and HDD separately Is not it an accurate experiment? I experimented with 9 servers I had before experimenting with VMs, but the results of SSDs and HDDs were similar, so I divided VMs into HDDs and SSDs and experimented. > Why is there no difference between kafka benchmark tests on SSD and HDD? > - > > Key: KAFKA-4971 > URL: https://issues.apache.org/jira/browse/KAFKA-4971 > Project: Kafka > Issue Type: Test >Affects Versions: 0.10.0.0 > Environment: Oracle VM VirtualBox > OS : CentOs 7 > Memory : 1G > Disk : 8GB >Reporter: Dasol Kim > > I installed OS and kafka in the two SSD and two HDDs to perform the kafka > benchmark test based on the disc difference. As expected, the SSD should show > faster results, but according to my experimental results, there is no big > difference between SSD and HDD. why? Ohter settings have been set to default. > *test settings > zookeeper node : 1, producer node : 2, broker node : 2(SSD 1, HDD 1) > test scenario : Two producers send messages to the broker and compare the > throughtput per second of kafka installed on SSD and kafka on HDD > command : ./bin/kafka-producer-perf-test.sh --num-records 100 > --record-size 2000 --topic test --throughput 10 --producer-props > bootstrap.servers=SN02:9092 > -- This message was sent by Atlassian JIRA (v6.3.15#6346)
[jira] [Commented] (KAFKA-4971) Why is there no difference between kafka benchmark tests on SSD and HDD?
[ https://issues.apache.org/jira/browse/KAFKA-4971?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15955260#comment-15955260 ] Michal Borowiecki commented on KAFKA-4971: -- I'd venture a guess that you are limited by something else than your hdd/ssd performance. Is 1g your total memory in the VM? How much of it is allocated to the kafka jvm process? Some things I can think of: Is there a lot of activity in the gc.log? Is the OS not swapping ferociously due to over-allocation of memory by any chance? Hope that helps. > Why is there no difference between kafka benchmark tests on SSD and HDD? > - > > Key: KAFKA-4971 > URL: https://issues.apache.org/jira/browse/KAFKA-4971 > Project: Kafka > Issue Type: Test >Affects Versions: 0.10.0.0 > Environment: Oracle VM VirtualBox > OS : CentOs 7 > Memory : 1G > Disk : 8GB >Reporter: Dasol Kim > > I installed OS and kafka in the two SSD and two HDDs to perform the kafka > benchmark test based on the disc difference. As expected, the SSD should show > faster results, but according to my experimental results, there is no big > difference between SSD and HDD. why? Ohter settings have been set to default. > *test settings > zookeeper node : 1, producer node : 2, broker node : 2(SSD 1, HDD 1) > test scenario : Two producers send messages to the broker and compare the > throughtput per second of kafka installed on SSD and kafka on HDD > command : ./bin/kafka-producer-perf-test.sh --num-records 100 > --record-size 2000 --topic test --throughput 10 --producer-props > bootstrap.servers=SN02:9092 > -- This message was sent by Atlassian JIRA (v6.3.15#6346)
[jira] [Commented] (KAFKA-4971) Why is there no difference between kafka benchmark tests on SSD and HDD?
[ https://issues.apache.org/jira/browse/KAFKA-4971?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15955219#comment-15955219 ] Dasol Kim commented on KAFKA-4971: -- I was experimenting with the kafka producer installed on the SSD by sending messages to one broker installed on another ssd server in throughput / sec. In the same way, the experiment was conducted using the kafka of the HDD installed server. According to the known facts, SSD experiments should show more throughput per second. test settings OS : CentOs 7, Memory : 1G, Disk : 8G it's my test result and command VM - SSD(producer 1, broker 1, partition 1) ./bin/kafka-producer-perf-test.sh --num-records 100 --record-size 2000 --topic test --throughput 10 --producer-props bootstrap.servers=SN02:9092 9169 records sent, 1833.8 records/sec (3.50 MB/sec), 2393.7 ms avg latency, 3621.0 max latency. 17840 records sent, 3566.6 records/sec (6.80 MB/sec), 4412.3 ms avg latency, 5282.0 max latency. 23568 records sent, 4704.2 records/sec (8.97 MB/sec), 4391.3 ms avg latency, 5132.0 max latency. 32872 records sent, 6574.4 records/sec (12.54 MB/sec), 2612.5 ms avg latency, 3055.0 max latency. 39352 records sent, 7870.4 records/sec (15.01 MB/sec), 2001.6 ms avg latency, 2644.0 max latency. 41168 records sent, 8228.7 records/sec (15.69 MB/sec), 2003.0 ms avg latency, 2585.0 max latency. 40568 records sent, 8105.5 records/sec (15.46 MB/sec), 2113.5 ms avg latency, 2374.0 max latency. 65528 records sent, 13105.6 records/sec (25.00 MB/sec), 1351.3 ms avg latency, 1778.0 max latency. 108096 records sent, 21619.2 records/sec (41.24 MB/sec), 780.2 ms avg latency, 1026.0 max latency. 79992 records sent, 15988.8 records/sec (30.50 MB/sec), 855.3 ms avg latency, 2238.0 max latency. 31152 records sent, 6230.4 records/sec (11.88 MB/sec), 2651.9 ms avg latency, 3180.0 max latency. 39520 records sent, 7899.3 records/sec (15.07 MB/sec), 1820.1 ms avg latency, 2536.0 max latency. 52824 records sent, 10564.8 records/sec (20.15 MB/sec), 1942.5 ms avg latency, 3243.0 max latency. 68912 records sent, 13760.4 records/sec (26.25 MB/sec), 1122.5 ms avg latency, 1678.0 max latency. 93024 records sent, 18601.1 records/sec (35.48 MB/sec), 954.2 ms avg latency, 1624.0 max latency. 80200 records sent, 16040.0 records/sec (30.59 MB/sec), 981.6 ms avg latency, 1417.0 max latency. 101720 records sent, 20344.0 records/sec (38.80 MB/sec), 829.8 ms avg latency, 1201.0 max latency. 73272 records sent, 14654.4 records/sec (27.95 MB/sec), 1076.5 ms avg latency, 1487.0 max latency. 100 records sent, 11094.223238 records/sec (21.16 MB/sec), 1444.94 ms avg latency, 5282.00 ms max latency, 1181 ms 50th, 3116 ms 95th, 4794 ms 99th, 5158 ms 99.9th. VM - HDD(producer 1, broker 1, partition 1) ./bin/kafka-producer-perf-test.sh --num-records 100 --record-size 2000 --topic test --throughput 10 --producer-props bootstrap.servers=SN03:9092 11145 records sent, 2228.6 records/sec (4.25 MB/sec), 2209.9 ms avg latency, 3442.0 max latency. 19592 records sent, 3912.9 records/sec (7.46 MB/sec), 4165.9 ms avg latency, 4661.0 max latency. 18472 records sent, 3694.4 records/sec (7.05 MB/sec), 4416.4 ms avg latency, 4598.0 max latency. 33312 records sent, 6662.4 records/sec (12.71 MB/sec), 2862.6 ms avg latency, 4366.0 max latency. 49000 records sent, 9782.4 records/sec (18.66 MB/sec), 1923.6 ms avg latency, 2673.0 max latency. 41856 records sent, 8357.8 records/sec (15.94 MB/sec), 1760.9 ms avg latency, 2241.0 max latency. 48032 records sent, 9602.6 records/sec (18.32 MB/sec), 1863.5 ms avg latency, 2283.0 max latency. 78032 records sent, 15606.4 records/sec (29.77 MB/sec), 1096.7 ms avg latency, 1364.0 max latency. 93440 records sent, 18688.0 records/sec (35.64 MB/sec), 833.6 ms avg latency, 1299.0 max latency. 72184 records sent, 14436.8 records/sec (27.54 MB/sec), 1185.9 ms avg latency, 1421.0 max latency. 80352 records sent, 16070.4 records/sec (30.65 MB/sec), 955.9 ms avg latency, 1896.0 max latency. 64200 records sent, 12840.0 records/sec (24.49 MB/sec), 1319.0 ms avg latency, 1652.0 max latency. 86400 records sent, 17280.0 records/sec (32.96 MB/sec), 972.3 ms avg latency, 1292.0 max latency. 74472 records sent, 14894.4 records/sec (28.41 MB/sec), 1073.2 ms avg latency, 1224.0 max latency. 75912 records sent, 15182.4 records/sec (28.96 MB/sec), 961.6 ms avg latency, 1901.0 max latency. 30088 records sent, 6017.6 records/sec (11.48 MB/sec), 2746.6 ms avg latency, 3482.0 max latency. 70368 records sent, 14073.6 records/sec (26.84 MB/sec), 1257.5 ms avg latency, 2300.0 max latency. 100 records sent, 11269.129347 records/sec (21.49 MB/sec), 1423.68 ms avg latency, 4661.00 ms max latency, 1179 ms 50th, 3467 ms 95th, 4460 ms 99th, 4597 ms 99.9th. Comparing the two experimental results, we can see that the results of the two experiments show little difference. I do not know what
[jira] [Commented] (KAFKA-4971) Why is there no difference between kafka benchmark tests on SSD and HDD?
[ https://issues.apache.org/jira/browse/KAFKA-4971?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15953608#comment-15953608 ] Michal Borowiecki commented on KAFKA-4971: -- I think your question would be easier to respond to if you quantified it by providing your test results and the drive specs. Kafka IO access patterns are designed to be sequential for good reason. Spinning disks and OS level buffering are optimised for such IO patterns, but I don't know if that alone can account for the miss-match between your expectations and the results your getting on your hardware. > Why is there no difference between kafka benchmark tests on SSD and HDD? > - > > Key: KAFKA-4971 > URL: https://issues.apache.org/jira/browse/KAFKA-4971 > Project: Kafka > Issue Type: Test >Affects Versions: 0.10.0.0 > Environment: Oracle VM VirtualBox > OS : CentOs 7 > Memory : 1G > Disk : 8GB >Reporter: Dasol Kim > > I installed OS and kafka in the two SSD and two HDDs to perform the kafka > benchmark test based on the disc difference. As expected, the SSD should show > faster results, but according to my experimental results, there is no big > difference between SSD and HDD. why? Ohter settings have been set to default. > *test settings > zookeeper node : 1, producer node : 2, broker node : 2(SSD 1, HDD 1) > test scenario : Two producers send messages to the broker and compare the > throughtput per second of kafka installed on SSD and kafka on HDD > command : ./bin/kafka-producer-perf-test.sh --num-records 100 > --record-size 2000 --topic test --throughput 10 --producer-props > bootstrap.servers=SN02:9092 > -- This message was sent by Atlassian JIRA (v6.3.15#6346)
