Experiment using a high CPU runner for assets compilation jobs

Related to #213107.

Duration of gitlab:assets:compile_webpack_if_needed per job / runner

Job / Runner	gsrm (n1-standard-2)	e2-standard-2	e2-standard-4	e2-standard-8	e2-standard-16	c2-standard-4	c2-standard-8	c2-standard-16
compile-assets pull-cache (1st run)	69s	151s	142s	126s	133s	84s	80s	77s
compile-assets pull-cache (2nd run)	86s	176s	142s	140s	136s	86s	80s	76s
gitlab:assets:compile pull-cache (1st run)	1296s	1417s	1002s	826s	754s	646s	534s	518s
gitlab:assets:compile pull-cache (2nd run)	1428s	1460s	964s	864s	774s	652s	562s	482s

Time gains

• Reference timing of gitlab:assets:compile_webpack_if_needed for the compile-assets pull-cache job: 77 seconds
• Reference timing of gitlab:assets:compile_webpack_if_needed for the gitlab:assets:compile pull-cache job: 1362 seconds

Job / Runner	gsrm (n1-standard-2)	e2-standard-2	e2-standard-4	e2-standard-8	e2-standard-16	c2-standard-4	c2-standard-8	c2-standard-16
compile-assets pull-cache	1	163 / 77 = 2.11	142 / 77 = 1.84	133 / 77 = 1.72	134 / 77 = 1.74	85 / 77 = 1.10	80 / 77 = 1.03	76 / 77 = 0.98
gitlab:assets:compile pull-cache	1	1438 / 1362 = 1.05	983 / 1362 = 0.72	845 / 1362 = 0.62	764 / 1362 = 0.56	649 / 1362 = 0.47	548 / 1362 = 0.40	500 / 1362 = 0.36

Prices per month - https://cloud.google.com/compute/vm-instance-pricing

N1 standard machine types (General-purpose machine type family) (current GitLab shared runners)

Machine type	Virtual CPUs	Memory	Price (USD)	Preemptible price (USD)
n1-standard-2	2	7.5GB	$48.5500	$14.60

E2 standard machine types (General-purpose machine type family)

Machine type	Virtual CPUs	Memory	Price (USD)	Preemptible price (USD)
e2-standard-2	2	8GB	$48.92	$14.67
e2-standard-4	4	16GB	$97.83	$29.35
e2-standard-8	8	32GB	$195.68	$58.7
e2-standard-16	16	64GB	$391.35	$117.41

C2 machine types (Compute-optimized machine type family)

Machine type	Virtual CPUs	Memory	Price (USD)	Preemptible price (USD)
c2-standard-4	4	16	$121.9392	$36.8650
c2-standard-8	8	32	$243.8784	$73.8030
c2-standard-16	16	64	$487.7568	$147.5330

Analysis: cost * time gain

Job / Runner	gsrm (n1-standard-2)	e2-standard-2	e2-standard-4	e2-standard-8	e2-standard-16	c2-standard-4	c2-standard-8	c2-standard-16
compile-assets pull-cache	48.5500 * 1 = 48.5500	48.92 * 2.11 = 103.22	97.83 * 1.84 = 180.00	195.68 * 1.72 = 336.56	391.35 * 1.74 = 680.949	121.9392 * 1.10 = 134.13	243.8784 * 1.03 = 251.19	487.7568 * 0.98 = 478.00
gitlab:assets:compile pull-cache	48.5500 * 1 = 48.5500	48.92 * 1.05 = 51.36	97.83 * 0.72 = 70.43	195.68 * 0.62 = 121.32	391.35 * 0.56 = 219.15	121.93 * 0.47 = 57.31	243.87 * 0.40 = 97.55	487.75 * 0.36 = 175.59

Conclusion

Based on the cost * time gain analysis, it looks like the c2-standard-4 machines would be have the best cost * time gain ratio (it costs 18% more than the n1-standard-2 for a 53% faster compilation). I suggest we use this type of machine only for the gitlab:assets:compile pull-cache job since there's almost no gain for the compile-assets pull-cache job.