What is the strategy for using Borg?

This article introduces the relevant knowledge of "what is the use strategy of Borg". In the operation of actual cases, many people will encounter such a dilemma. Then let the editor lead you to learn how to deal with these situations. I hope you can read it carefully and be able to achieve something!

Use efficiency

One of the main purposes of Borg is to make effective use of Google's machine fleet, which is a large financial investment: millions of dollars can be saved by increasing efficiency by a few percentage points.

1 measurement methodology

Our job deployment is resource-constrained, and we rarely encounter peak loads, our machines are heterogeneous, and the resources we recycle from service job run batch job. Therefore, in order to measure, we need a more abstract standard than "average utilization". After doing some experiments, we chose cell density (cell compaction): given a load, we kept starting from scratch (to avoid being stuck by an unlucky configuration) and deployed to the smallest possible Cell until we could no longer extract the machine from the cell. This provides a clear termination condition and facilitates automated comparison without traps, which refer to integrated workloads and modeling [31]. A quantitative comparison and estimation technique can be seen in [78], with many subtle details.

We can't do performance experiments with online cell, so we use Fauxmaster to achieve high-fidelity simulations, using real online cell load data including all constraints, actual limitations, reservations, and commonly used data ($5.50). This data is extracted from the Borg snapshot (checkpoints) of PDT at 14:00 on 2014-10-1. (other snapshots produce similar conclusions). We selected 15 Borg cell to report, excluding special purpose, test, and small (