Comment by tayo42
1 day ago
If you have 50,000 servers for your service, and you can reduce that by 1 percent, you save 50 servers. Multiply that by maybe $8k per server and you have saved $400k,you just paid for your self for a year. With meta the numbers are probably a bit bigger.
yes, but latency-optimized schedulers tend to have _worse_ throughput, not better.
LOL (I used to work for Meta, so appreciate the facetious understatement)
That's not how it works though. Budgets are annual. A 1% savings of cpu cycles doesn't show up anywhere, it's a rounding error. They don't have a guy that pulls the servers and sells them ahead of the projection. You bought them for 5 years and they're staying. 5 years from now, that 1% got eaten up by other shit.
You're wrong about how services that cost 9+ figures to run annually are budgeted. 1% CPU is absolutely massive and well measured and accounted for in these systems.
So you prematurely dump hardware you already own when you see CPU usage go down? I don't think so.
1 reply →
You don't buy servers once every 5 years. I've done purchasing every quarter and forecasted a year out. You reduce your services budget for hardware by the amount saved for that year.
5 years is the lifecycle. You're not going to get rid of a 4 year old server because you're using less cycles that you thought you would. You already bought it. You find something else for it to do or you have a little extra redundancy. If I increase the mpg of my semi fleet, that doesn't mean I can sell some of my semis off just because the cost per trip goes down.