These are just some observations from poking around the data a couple minutes. Now if only EC2 would allow bidding across a range of instance types to run your jobs, folks could get more done with less. Will someone manipulating the market price, pushing it beyond reasonable viability?.Can you run your AMI on a different instance type?.Do you really need your spot in the same AZ or region?.Do you need an on-demand instance when a spot will do?.ConclusionsĪWS originally marketed the spot market as allowing you to run work at different times for more optimal pricing. But, you should also ask yourself: For your shillings, you’d likely be better with a c3.large than a m1.medium, but the market data does not indicate consumers realize this as the price for c3.large instances is a flatline. That’s odd, given this is a compute node, of which you’d expect to see large numbers being recruited and released for brief computation cluster bursts. This makes the average clearing price for a c3.large ($0.032/hr) 97% lower than an m1.medium ($1.26). For example, why is the c3.large instance, which is far better than a m1.large instance, always available for $0.032/hr? Over the last three months, there is not a spike to be seen. In the 3 months of data analyzed we see notable spot price disparities between neighboring zones. Sure, some jobs might need very high intra-AZ networking speeds, but most can (and should) be spread across AZ’s. If the price for work in one AZ is much higher than another, near by AZ, we’d expect to see work migrate there. Highly different demand between instance types ( c3.large low, m1.medium high).Highly different demand between AZ’s (i.e. However, demand for compute is oddly inconsistent across regions, zones, and instance type. If the price for beer in a 1 liter container is up, we’d expect to see a similar rise in demand for beer in the 2 liter container. Other Observations – Highly Non-Correlated Market Demand Regardless, it renders much of the market useless and highly inefficient. They don’t want the spot pool available, perhaps for maintenance reasons or some internal load shaping. The only explanation is that during these hours Amazon (or someone with money to burn) is shutting down the market. You don’t see people winning an auction by 5000% do you? You pay $0.01+ and call it a victory. Also note that the price is always going to $6 in a step function, while in a real market, with sentient traders, we’d expect to just barely outbid our competitors as anything else is waste. That means the entire market is going for $6. Perhaps one foolhardy API consumer put in too high a bid, but we’re talking a $6 market clearing price. So, what gives? Are there really people that are so gormless they shovel money to Amazon? Impossible. $6/hr is 30% more than that most expensive instance AWS sells ( hs1.8xlarge, $4.60/hr), and this instance is puny by comparison.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |