These tests evaluate the performance of single and multiple MDS servers when handling requests from multiple clients. These clients are building the kernel: first unpacking the source, compiling the kernel, and finally removing all source and compiled files.
These tests are run using ceph-ansible on Linode virtual machines. Each test uses 3 monitors, 3 MDSs (2 standby during single MDS tests), 16 OSDs, and 8 or 16 clients. All VMs use the basic 2048MB plan on Linode (i.e. 2GB RAM, 1 core). The one exception is tests using the 4096MB MDS servers (4GB RAM, 2 cores).
The time to execute the parallel build for the clients is given in the table below:
| RAM/Cores | # of Clients | # of MDSs | Time | Average Sum of Client Latency |
|---|---|---|---|---|
| 2GB/1 | 8 | 1 | 99m | 3326.75 |
| 2GB/1 | 8 | 2 | 80m | 1364.04 |
| 2GB/1 | 8 | 3 | 63m | 1307.50 |
| 2GB/1 | 16 | 1 | 323m | 16633.70 |
| 2GB/1 | 16 | 2 | 263m | (lost: corrupt db) |
| 2GB/1 | 16 | 3 | 84m | 2499.43 |
| 4GB/2 | 8 | 1 | 68m | 2037.89 |
| 4GB/2 | 8 | 2 | 73m | 1235.29 |
| 4GB/2 | 8 | 3 | 50m | 847.45 |
| 4GB/2 | 16 | 1 | 147m | 5277.31 |
| 4GB/2 | 16 | 2 | 115m | 3274.02 |
| 4GB/2 | 16 | 3 | 108m | 2876.89 |
When using a single core, the single MDS would consistently be at 100% CPU for the duration of the test for all configurations. With 2 cores, the single MDS would use approximately 150% CPU for the duration of the test. (Numbers about CPU usage were not collected. I observed this using htop.)
A full list of configurations and some graphs of performance metrics are here.