查看文章

Over the last year, a group of us at MIT, Stanford, CMU, Google, and VMware have been designing and implementing a new Operating System (OS) stack for modern hyperscale datacenter environments. This new stack leverages a set of multi-core, multi-node distributed DBMSs near the bottom of the stack to manage a cluster of machines on a public or private cloud. In this paper, we briefly review the rationale for a new OS, present our resulting architecture, and review our progress to date. The meat of our paper is a presentation of the main lessons thus far from this project. Many of these have to do with missing capabilities in multi-node DBMSs that form the guts of our proposal. Finally, we present future research directions in database-oriented operating systems.

1 Why a New OS? Current OSs like Linux are based on the UNIX design from the 1970s. Back then, hardware resources consisted of a uniprocessor with limited main memory and disk. In the intervening 40 years, resources under OS management have increased by five or six orders of magnitude. The MIT Supercloud [5], for example, has approximately 10,000 cores and a hundred terabytes of main memory in total. OS state (files, tasks, messages, etc.) has increased in size by the same scale factor. Hence, today’s massive scale systems are very different from what Linux was designed for. Over the past year, we have worked on designing a new OS stack for distributed environments called DBOS. We chose to base our stack on DBMS technology as it is known to provide the scale, performance, and reliability required in large-scale systems. For example, increasing network speeds …