Wire-once, provision-many: optimizing compute resources in the data center - Storage Networking

Computer Technology Review, May, 2003 by David Caplan

Data center architecture has taken strides toward logical organization in the past few years as the computing model has moved toward disaggregation. By breaking out the functions of storage, computing, and networking, and optimizing the delivery of each, enterprises have made some significant improvements in data center performance. Disaggregation, in general, provides three primary benefits:

Better Performance: When a function (storage, computing, or networking) is bundled together with other functions, the critical services are typically delivered by running software on general-purpose processors. This delivers far slower performance than is possible using purpose-built hardware devices. By isolating each function, thus limiting the scope of what each box must do, it becomes possible to deliver dedicated hardware and software that dramatically increases performance.

Superior Scalability: Because functions are isolated from each other, disaggregation gives data center operators the freedom to alter one function without impacting the other two. For example, storage capacity can be increased by adding more storage devices, without having to do anything to the computing and networking tiers. This point may seem intuitive, but it was not that long ago that it was a regular practice to increase storage capacity by buying more computers.

Improved Economics: Storage, computing and networking infrastructure can be deployed cost-effectively because provisioning is improved and disparate equipment is consolidated. Much of the complexity of administering each function is hidden, since each is isolated from the other. Disaggregation is also very tightly linked with virtualization, which allows for sharing of physical equipment across multiple customers, departments or applications. Virtualization significantly decreases up-front capital expense, as well as long-run operational expense, because there are simply fewer boxes to buy and manage, and compute resources can be managed as a consolidated pool, rather than on a box-by-box basis.

Disaggregation thus represents a major force in the evolution of the computing model in the data center. The storage industry has led the charge into disaggregation, with virtualized storage, and has made significant progress in proving the benefits of this approach. Virtualized storage has enabled data center operators to make far more efficient use of their storage resources by managing them as a single shared pool of resources, independent of the computing and networking tiers. This eliminates the phenomenon of one storage device being jammed to capacity while another sits idle.

However, while virtualization has improved storage efficiency, it is a different story m the computing tier. The computing tier of data centers still looks like a collection of hardwired boxes, each dedicated to a specific application silo. This creates a situation where you'll see, for example, the machines running PeopleSoft wheezing at maximum capacity, while the 12 application across the way is barely being used. Wouldn't it make sense if the PeopleSoft application could "borrow" computing power from the boxes running the 12 application? Of course it would, but there's never been a way to dynamically provision computing resources to make this happen.

Today, this is changing. There is a new trend afoot: Virtualized computes. Just like virtualized storage, virtualized computes lets data center operators manage the computing tier as a single, shared pool of resources, where computing power can be provisioned on the fly to optimize the performance of applications. This enables both the simplification and optimization of the computing tier, resulting in a more efficient and productive data center.

Virtualizing Computes in the Data Center

There are a number of basic requirements to successfully virtualizing computing power in the data center. First, there must be a way to offload compute-intensive overhead like SSL processing and TCP termination, so it does not bog down the compute resources. SSL decryption is also critical because there must be layer 4-7 switching in place, which can make intelligent decisions on resource allocation and protect servers from malicious application-level attacks. Obviously this cannot be accomplished if the traffic is encrypted.

Furthermore, it's one thing to inspect incoming traffic and provision the appropriate computing power to service it. It's another thing to do this at the gigabit speeds required to maintain appropriate levels of performance. So the devices performing all these functions must be hardware-based.

Traditional switches may scale effectively, but given their packet-centric architecture, they are not designed to deliver higher layer application services like protection against Nimda and CodeRed. They can't inspect incoming data up to layer 7 and still sustain gigabit scale. Software-based appliances are more application-aware than switches, but fall short in the throughput department.