 Linux Kernel Performance |
| LTC Home | IBM Linux Scalability Effort | LinuxKernelPerformance Project | Performance Tools | Disk I/O | IOZONE | LMBench | Netbench | Netperf | SPEC SDET | SPECWeb99 | tiobench | VolanoMark |
Linux Kernel Performance |
| LTC Home | IBM Linux Scalability Effort | LinuxKernelPerformance Project | Performance Tools | Disk I/O | IOZONE | LMBench | Netbench | Netperf | SPEC SDET | SPECWeb99 | tiobench | VolanoMark |
VolanoMarkTM 2.1.2 is a benchmark of a chat room server that is written in the JavaTM language. The benchmark was developed by Volano, LLC to compare their chat server performance under different Java implementations and operating systems. The performance metric reported by a VolanoMark test is "chat server" messages per second. We use Volanomark to measure and analyze the performance of the linux kernel, particularly the scheduler, signals, tcp/ip stack, and the user space pthread library. There is a client and server side of the benchmark so it can be run in a loopback or network mode.
Volano, LLC publishes The Volano Report comparing VolanoMark performance of several Java implementations and operating systems. Volano, LLC runs both a loopback and network test. The May 29 2001 Volano Report presents VolanoMark loopback test results that report the IBM JVM 1.3.0 (20010502) on Red Hat Linux release 7.1 using the 2.4.2-2 kernel outperforms all other Java implementations and operating systems tested. The same report also presents results for the VolanoMark network test where this same JVM on linux under performed most of the other Java implementations and operating systems tested. The VolanoMark network test, as run by Volano, is a measure of how many network connections can be created when using a particular Java implementation and OS. The under performance in the network test will be one area of our investigation.
Status
May 1, 2002 - VolanoMark benchmark results for UP improved 33.6 % from 14,986 to 20,026 msg/sec. which is 160.2 % of the UP throughput target. 8P results improved 24.8 % from 50,994 to 63,648 msg/sec which is 84.9 % of the 8P throughput target.
Results
| Coming | Scheduler statistics data for a VolanoMark 10/100 loopback test of the 2.4.4 kernel. |
| Soon... | Explanation of what schedstat reveals for this workload. |
| IBM JVM 122 (20001026), 2.4.4 kernel, 8 CPUs, schedstat2_0_0-2_4_4.patch, schedstat2_0_0.tar.gz | |
| Coming | VolanoMark 10/100 loopback test results comparing the IBM JVM 1.2.2 and IBM JVM 1.3 on a 2.4.4 kernel. |
| Soon... | IBM JVM 1.2.2 (20001026), IBM JVM 1.3 (20000502), 2.4.4 kernel, 8 CPUs |
| 14 Jun 2001 | Lockmeter data for a VolanoMark 10/100 loopback test of the 2.4.4 and 2.4.4+MQ Scheduler kernel. |
| Using the 2.4.4 MQ scheduler patch on an 8P reduced utilization and spin time for the runqueue_lock by 4 CPUs. | |
| IBM JVM 122 (20001026), 2.4.4 kernel, 8 CPUs, 2.4.4.mq1-sched patch, -Xms64MB -Xmx64MB | |
| 04 Jun 2001 | Kernprof acg data for a VolanoMark 10/100 loopback test of the 2.4.4 and 2.4.4+MQ Scheduler kernel. |
| IBM JVM 122 (20001026), 2.4.4 kernel, 8 CPUs, 2.4.4.mq1-sched patch, -Xms64MB -Xmx64MB | |
| 19 May 2001 | VolanoMark 10/100 loopback test results for 2.4.4+MQ kernel comparing 8MB and 64MB initial JVM heap size. |
| Using a 64 MB initial JVM heap size improved 8P throughput by 27 % and 4P by 17% for the 2.4.4+MQ kernel. | |
| IBM JVM 122 (20001026), 2.4.4 kernel, 8 CPUs, 2.4.4.mq1-sched patch | |
| 19 May 2001 | VolanoMark 10/100 loopback test results for affinitizing the Volano server and client to separate CPUs. |
| Process affinity improved 8P throughput by 25 % for the 2.4.4 kernel. | |
| IBM JVM 122 (20001026), 2.4.4 kernel, 8 CPUs, prctl_2_4_0.patch, prctlaff.c, -Xms64MB -Xmx64MB. | |
| 19 May 2001 | VolanoMark 10/100 loopback test results for the 2.4.4 and 2.4.4+MQ scheduler kernel. |
| Using the MQ scheduler patch for the 2.4.4 kernel improved 8P throughput by 147 % and 4P by 32 % | |
| IBM JVM 122 (20001026), 2.4.4 kernel, 8 CPUs, 2.4.4.mq1-sched patch, -Xms64MB -Xmx64MB | |
| 19 May 2001 | VolanoMark 10/100 loopback test results for 2.4.4 kernel comparing 8MB and 64MB initial JVM heap sizes. |
| Increasing the JVM initial heap size from 8MB to 64MB improved 8P throughput by 10 % and 4P by 13%. | |
| IBM JVM 122 (20001026), 2.4.4 kernel, 8 CPUs | |
| 25 Apr 2001 | VolanoMark 10/100 loopback test results for 2.4.1 kernel comparing 8MB and 64MB initial JVM heap sizes. |
| Increasing the JVM initial heap size from 8MB to 64MB improved 8P throughput by 12% and reduced idle time. | |
| IBM JVM 122 (20001026), 2.4.1 kernel, 8 CPUs, 2.4.1.mq1-sched patch. | |
| 24 Apr 2001 | Lockmeter data for a VolanoMark 10/100 loopback test of the 2.4.1+MQ Scheduler kernel. |
| IBM JVM 122 (20001026), 2.4.1 kernel, 8 CPUs, 2.4.1.mq1-sched patch, -Xms8MB -Xmx64MB | |
| 24 Apr 2001 | Kernprof acg data for a VolanoMark 10/100 loopback test of the 2.4.1+MQ Scheduler kernel. |
| schedule() called often - many sys_sched_yield(), signals (pthread mutex locks), and tcp_data_wait | |
| IBM JVM 122 (20001026), 2.4.1 kernel, 8 CPUs, 2.4.1.mq1-sched patch, -Xms8MB -Xmx64MB | |
| 03 Apr 2001 | VolanoMark 10/100 loopback test results for the 2.4.1 and 2.4.1+MQ scheduler kernel. |
| Using the MQ scheduler patch for the 2.4.1 kernel improved 8P throughput by 150 % and 4P by 37 %. | |
| IBM JVM 122 (20001026), 2.4.1 kernel, 8 CPUs, 2.4.1.mq1-sched patch, -Xms8MB -Xmx64MB | |
| 19 Feb 2001 | VolanoMark 10/100 loopback test results of the SCHED_YIELD patch for 2.4.0-test9 and 2.4.0.test10 kernel. |
| Fixing the SCHED_YIELD bug and the runqueue_lock optimizations improved 8P throughput by 72 %. | |
| IBM JVM 122 (20001026), 2.4.0-test8/9/10, 2.4.0 kernel, 8 CPU, -Xms8MB -Xmx64MB |
Test Environment
Resources
Bill Hartner hartner@austin.ibm.com
Partha Narayanan partha@us.ibm.com
N/A
The run rules and H/W we use are different than the run rules and H/W used by Volano, LLC, so the results can not always be compared with those presented in The Volano Report.
N/A