Afterwards I had a look at the optimal division between reserved and spot instances, but without taking into account more properties of using spot instances (such as the need for checkpointing, ...). Using spot instances seemed to be cheaper than using reserved instances, such that the optimal division is using a 100 percent spot instances.
During the thesis meeting last week another question was raised: whether it is cheaper to under- or overestimate the number of required Reserved instances to reach an optimal division? While trying some different workloads in the earlier created Excel sheet, I found out this (as one would expect) depends on the workload. I tried to find a formula that expressed whether for a certain workload it would be better to take one less or one more reserved instance. Of course this is a bit contradictious, since you could determine the optimal amount of reserved instances if the workload is known. But maybe the found tipping point could tell us something, my findings can be found here.
No comments:
Post a Comment