Scalability of Linux kernel 2.4 and kernel 2.6

The workload for this test was Dbench version 2.1. The benchmark simulates a file server and belongs to the benchmark family of Tbench, Dbench, Smbtorture. Dbench measures only the server without network. The load generator is on the same system. A single process of the load generator drives a predefined set of file system operations. The elapse time of all operations is measured and an average throughput is calculated by dividing the total file size of all processed files by the elapse time. Starting several workload processes (clients) allows workload scaling. For all tests the memory size was 2GB. The workload was tested with different numbers of CPUs (1-4-8-16).

The scaling factors on SLES8 are 1.7 (2 CPUs), 3.3 (4 CPUs), 5.0 (8 CPUs) and 5.4 (16 CPUs). The same tests on SLES9 showed a similar throughput for 1, 2 and 4 CPUs. The 8 and 16 CPU tests performed much better. The CPU scaling factors on SLES9 are 2.0 (2 CPUs), 3.8 (4 CPUs), 7.2 (8 CPUs) and 12.7 (16 CPUs).

The promised scaling improvement in kernel 2.6 when using more CPUs can be seen here.

Back to top

A second scaling test was made with an Informix 64bit database on SLES8 (z900t) and SLES9 (z990). The purpose was not to run a typical database benchmark with huge memory and storage server requirements but rather to test CPU and process scalability. As workload we used transactions simulating a complete environment in which a population of terminal operators performs transactions against the database. The benchmark is centered around the principal activities (transactions) of an order-entry environment.

The scaling in the SLES8 tests show scaling factors of 3.7 (4 CPUs), 6.9 (8 CPUs), 9.5 (12 CPUs) and 11.5 (16 CPUs). SLES9 on z990 shows a similar scaling, the throughput was 50% higher than with SLES8 and z900t when more than 1 CPU was used in the test.

Back to top