Plastic brain learning/un-learning

Archive for the ‘Energy efficient’ Category

Path to greener Data Centers, Clouds and beyond….

In Cloud Computing, Data center, Energy efficient on December 7, 2008 at 7:49 am

Data center power consumption (w.r.t IT equipment like Servers, Storage) is a hot topic these days. The cumulative install base of Servers around the globe is estimated to be in the range of 50 million units, and in the grand scheme of things, might consume 1% of the world-wide energy consumption (according to estimates from Google, let me know if you need a reference). If you’re wondering, yes, world-wide install base of PC’s might consume several multiples of Server power consumption in Data center….why? Because world-wide PC install base is upwards of 600 million units!

Data center power consumption obviously gets more attention…because of the concentration of power consumption. 50 Mega Watts of power consumption at one Data center location is more conspicuous than the same amount of power consumed by PC’s in millions of households and businesses.

Trying to understand the trends towards more eco-friendly computing requires understanding of developments at many levels. Starting from the VLSI design innovations at the Chip or Processor level, Board level, Software level (Firmware, OS/Virtualization, Middleware, Application level), and finally at the Data center level.

Chip or Processor level: Processor chips are already designed to work at a lower frequency based on load, in addition to providing support for virtualization. In Multi-core chips, cores can be off-lined, depending on need (or problems). Chips are designed with multiple power domains….so CPU’s can draw less power based on utilization. The issue is with other parts of the computer system such as Memory, Disks etc. Can you ask the memory chips to offline pages and draw less power? Can you distribute data across Flash or Disks optimally to allow similar proportional power consumption based on utilization levels? These are certainly some of the dominant design issues that need to be addressed, keeping in mind constraints such as low-latency, little or no “wake-up” penalty.

Board level: Today, Server virtualization falls short of end-end virtualization. When machine resources are carved up, guest VM’s don’t necessarily carve up the hardware resources in proportion. Network level virtualization is just beginning to evolve. For e.g. Crossbow in OpenSolaris. Another example is  Intel’s VT technology: enables allocation of specific I/O resources (graphics card or network interface card) to guest VM instances. If Chips and Board level hardware elements are power (and virtualization) savvy, you can ensure power consumption that is (almost) proportional to utilization levels, dictated by the workload.

Firmware level: Hypervisors, whether Emulated or Para-virtualized, present a single interface to the hardware, and can exploit all the Chip-level or board-level support for “proportional energy use” against a given workload.

OS level: Over a sufficiently long time interval (months), server utilization is predominantly characterized by low utilization intervals. Average utilization of Servers in Data centers is usually less than 50%. That means there is plenty of opportunity for Servers to go in to “low-power” mode. How can you design the OS to co-operate here? 

System level: Manageability (e.g. responding to workload changes, migrating workload seamlessly etc), Observability (e.g DTrace ), API’s to manage Middleware or Application stack in response to low-power mode of operation (again, proportional power usage w.r.t workload) are going to be paramount considerations.

Cloud level: shouldn’t Clouds look like operating systems (seamless storage, networking, backups, replication, migration of apps/data, dependencies similar to pkg dependecies). 3Tera and Rightscale solve only some of these problems…but many areas need to be addressed: Dynamic, workload based Performance qualification, Mapping application criticality to Cloud Deployment Models, Leveraging Virtualization technologies seamlessly…

Data center level: Several innovations outside of IT (HVAC systems, again enabled by IT/sensor technologies), as well as innovations at all of the levels discussed above will help drive down PUE (Power Usage Effectiveness) at the Data center level closer to the holy grail (PUE = 1, i.e all the energy supplied to the Data Center goes to useful compute work done). Microsoft’s Generation4 effort represents a leap in this domain, as more and more companies realize that this is a big change of paradigm, as computing business truly goes in to utility scale/mode.

So, there are plenty of problems to be solved in the IT space….pick yours at any of these levels 🙂