extend docs for benchbase chbenchmark#178
Conversation
|
|
||
| The benchmark config file specifies the workload parameters. | ||
| The following file specifies the CH workload with scale factor 100, which is about 10 GB data. | ||
|
|
There was a problem hiding this comment.
We could include docker setup infos here or above, especially the command line args to set the postgres user password accordingly
There was a problem hiding this comment.
We also should decide if we want to enable async mode (was enabled for below). This does not matter for local machines, but for EBS with fsync latency around 4ms and writethrough reporting it does
| <warmup>60</warmup> | ||
| <time>120</time> |
There was a problem hiding this comment.
If we don't expect a clean create -> load -> execute run we should have a longer warmup so slow disks can actually read all data during warmup
| <warmup>60</warmup> | |
| <time>120</time> | |
| <warmup>180</warmup> | |
| <time>300</time> |
| ### Results | ||
|
|
||
| After the benchmark run, benchbase prints a detailed report of the workload. | ||
| The following is an example run on AWS with EBS. TODO: @ChrisWint please add the specifics. |
There was a problem hiding this comment.
Measurements below are on https://instances.vantage.sh/aws/ec2/m7a.16xlarge with a 1GB/s 16k IOPS gp3 volume, but for scale factor 1000 with 100 terminals. We might want to remeasure according to the xml we have above to give accurate results.
There was a problem hiding this comment.
Might want to have a (rough) sizing guide per scale factor. For best performance instance memory should be 300mb per warehouse to be on the safe side regarding memory handling. wrt CPU cores I don't know exactly when we start to suffer from the Linux scheduler under oversubscription, but I think 10 connections/terminals per core should be safe still. Lower doesn't hurt either
|
|
||
| ```text | ||
| 67 queries / 120s * 3600 = 2.010 QphH | ||
| ``` |
There was a problem hiding this comment.
Do we want to include our chbench tool here as well? I don't think it is necessary for this PR, just raising the possibility
2 participants
No description provided.