1E’s fellow co-founder and all round geek Mark Blackburn presented at the Green Grid Technical Forum this week. His chosen topic of expertise was the upcoming Data Center compute Efficiency (DCcE) metric.
Mark’s work in this field is really exciting as it moves data center efficiency onto an area which has been long neglected, that of Server Power Management. Historically data canter managers have been all about reducing cooling load, water usage and general data center power usage at the rack level. The DCcE metric allows you to go that extra step into the rack to determine the efficiency of the servers themselves.
Stepping back a bit, DCcE relies on the calculation of the efficiency at the individual server level. This is know as Server Compute Efficiency (SCE). SCE is a method of determining whether or not a server is performing its Primary Service either efficiently or even at all. So a server could be allocated as a Database host, and that would be the Primary Service for that server. Alongside the Primary Service however there will also be secondary and Tertiary services such as defragmentation, backup, virus scanning etc which will obviously cause some utilization of the machine.
So although Primary Service usage is the one to watch, in a data center of hundreds or maybe even thousands of servers, it is in fact much easier to track the utilization of those common secondary and tertiary services. If you do that, then you can end up with an accurate idea of your Primary Services utilization by simply working out the following: Primary service work = All work –Secondary & Tertiary work. Simple!
Or in it’s slightly more complex form, as Mark demonstrated this week, the following rules apply..
Over a time period
–All CPU minus secondary & tertiary CPU > noise threshold
–All I/O minus secondary & tertiary I/O > noise threshold
–There have been incoming network sessions for primary services
–There has been an interactive logon
–Server was being useful
–Server was not being useful
Having performed these calculations and measurements, we eventually end up with the Server Compute Efficiency (ScE), which is the proportion of samples that the server is providing Primary Services over time (as a percentage). This instantly provides us with the ability to detect unused servers, as any server with an ScE of 0% over a prolonged period of time must be doing nothing useful at all! Also, any servers with low ScE are worth investigation as they may be candidates for virtualization.
So when we have the ScE for all of our servers, we can determine the Data Center compute Efficiency (DCcE) by aggregating ScE across all servers in the data center. So here are some important facts worth noting about DCcE.
DCcE Provides a benchmark against which to improve (like PUE)
DCcE is NOT a productivity metric in that it does not measure how MUCH work is done, just the proportion of work that is useful
DCcE CANNOT and SHOULD NOT be compared between data centers due to the subjective determination of secondary and tertiary processes
DCcE is another great tool in the armory that is Data Center Efficiency, and opens up many new possibilities in this field.
Many thanks to Mark Blackburn for his great work on this. There will be a Green Grid whitepaper to follow, and if you want to see the slides from the whole presentation you can grab them here.
Also, if you are attending MMS2011 you can catch our session on Tuesday, 3:30 – 3:50 Datacenter Efficiency using NightWatchman Server Edition, as we use the very same DCcE calculations in our server power management product.