Benefits of using CWQs¶
The following is a good example of the benefits when implementing cluster wide quotas. The example makes use of Tekton.
Without Cluster Wide Quotas¶
Without Cluster Wide Quotas, every Tekton namespace would be linked to a specific cluster quota. All of these quota would add up to a considerable total quota reservation.
However, since Tekton pipelines only rarely actually use these quotas on average about 99% of quota would always be unused. The cluster would either be considerably over committed or underutilized.
With Cluster Wide Quotas¶
By enabling Cluster Wide Quotas, all Tekton namespaces will be linked to one cluster wide quota.
The Cluster Wide Quota can be automatically tuned by the Paas operator, having a minimum, maximum, default, and ratio value.
The algorithm comes down to using the largest value out of 3 different calculations:
- ratio (Paas config) * sum of all Paas resources (either defined or default);
- sum of the largest two of all Paas quotas (either defined or defaulted);
- a hard set minimum (Paas config);
Furthermore, a maximum value (Paas config) can be set to cap the quota.
Example with numbers¶
Consider the following values being set in the Paas config (e.a. for cpu.limits):
| cluster wide config | ||
|---|---|---|
| min value | 10 | When set, every cluster wide quota has at least this amount. |
| max value | 20 | When set, every cluster wide quota has at most this amount. |
| default | 5 | Resources can be set in Paas. When not set in Paas, this default is used. |
| ratio | 0,3 | Scales the sum of all quotas (all Paas'es) down to a lower value for cluster wide quota. |
Now imagine that we would have a cluster where one by one new Paas'es would be created, with Tekton enabled, all using the default quota config.
Quota reservations would be along lines of:
| # of PAASes | CWQ disabled | Scale by ratio | Scale with min values | Scale with hardset min | Scale with min & max | Savings |
|---|---|---|---|---|---|---|
| 1 | 5 | 1,5 | 10,0 | 13,0 | 13,0 | -8,0 |
| 2 | 10 | 3,0 | 10,0 | 13,0 | 13,0 | -3,0 |
| 3 | 15 | 4,5 | 10,0 | 13,0 | 13,0 | 2,0 |
| 4 | 20 | 6,0 | 10,0 | 13,0 | 13,0 | 7,0 |
| 5 | 25 | 7,5 | 10,0 | 13,0 | 13,0 | 12,0 |
| 6 | 30 | 9,0 | 10,0 | 13,0 | 13,0 | 17,0 |
| 7 | 35 | 10,5 | 10,5 | 13,0 | 13,0 | 22,0 |
| 8 | 40 | 12,0 | 12,0 | 13,0 | 13,0 | 27,0 |
| 9 | 45 | 13,5 | 13,5 | 13,5 | 13,5 | 31,5 |
| 10 | 50 | 15,0 | 15,0 | 15,0 | 15,0 | 35,0 |
| 11 | 55 | 16,5 | 16,5 | 16,5 | 16,5 | 38,5 |
| 12 | 60 | 18,0 | 18,0 | 18,0 | 18,0 | 42,0 |
| 13 | 65 | 19,5 | 19,5 | 19,5 | 19,5 | 45,5 |
| 14 | 70 | 21,0 | 21,0 | 21,0 | 20,0 | 50,0 |
| 15 | 75 | 22,5 | 22,5 | 22,5 | 20,0 | 55,0 |
| 16 | 80 | 24,0 | 24,0 | 24,0 | 20,0 | 60,0 |
| 17 | 85 | 25,5 | 25,5 | 25,5 | 20,0 | 65,0 |
| 18 | 90 | 27,0 | 27,0 | 27,0 | 20,0 | 70,0 |
| 19 | 95 | 28,5 | 28,5 | 28,5 | 20,0 | 75,0 |
Visualized in a graph, it would look like this:
Once 10 Paas'es would be created:
- without CWQ's, about 50 CPU limits would be reserved for all Tekton namespaces;
- with a ratio of 0.3 only 15 would be reserved, saving 35 CPU limits;
Beyond 14 Paas'es, there would only be 20 CPU's reserved no matter how much Paas'es would be created.