I love the analogy made in this paper: "Much like the first computers, the first datacenters ran a small number of applications written by expert users." This paper argues that datacenters need an operating system. Since the datacenter is being regarded as a computer itself, not just a collection of individual computers, their argument is sound and compelling.
The authors explains four roles required for the new OS: resource sharing, data sharing, programming abstractions, and debugging and monitoring, each of which has been a critical for traditional operating systems as well. As the authors admit, there have been significant efforts for each goal. What I am wondering is that what "putting it all together" can add more values than "loosely-coupled" datacenter middleware.
In the section 3, the slogan "Focus on paradigms, not performance" is thought-provoking. On the one hand, it is doubtless. For example, the very first file system of UNIX could utilize only 2% of the bandwidth of HDDs at that time. But its clever design turned out to be very useful in the test of time. For example, the abstraction of VFS could seamlessly support TCP/IP sockets, which have enabled today's Internet. On the other hand, performance and datacenter OS cannot go apart since the performance is the raison d'être of the datacenter applications.
No comments:
Post a Comment