As I wrote in the previous two blogs on energy optimization in manufacturing, Data Center Energy optimization faces exactly the same set of challenges. How do we know concurrent usage of machine and energy? When a fan is running on a server, is that cooling really necessary? In DCIM (Data Center Infrastructure Management software), power data is available on each PDU (power distribution unit) which typically powers up a cluster of servers or switches. To do optimization of local fans or global HVAC air cooler, against the usage of the servers, concurrent time series data of fan load vs usage for every server is required in addition to environmental data of ambient temperature in each section of PDU.
DCIM has come a long way in the last five years but they are not yet at a point where it can capture energy usage vs fan usage vs server load for each server. The reason is simple. They are capturing energy of a PDU via a smart meter and it can be only pitted against an aggregated cluster where each server may not have the same usage pattern. So PDU energy data is only good for environmental optimization of air cooler of HVAC.
To capture energy usage vs fan usage vs server load, one needs a different power monitoring strategy where using current voltage and power factor signature, one can extract all of those data concurrently. Consequently, this requires more machine learning and edge analytic but at the end, when thousands of servers are involved, and hundreds of them are added on a regular basis to cope up with future demand, one smart device that generates energy, fan usage and load from same point, will provide the granularity and concurrency of the data required for optimization.
As noted by a few experts, in today's data center PUE (ratio of input power to power used for computation, storing & switching) between 1.1-1.3 is considered a great job as opposed to 1.3-1.5 in the past.
Of course, servers are getting smarter, software defined PDU is evolving. IBM's System X, Dell's 12th-generation PowerEdge, are good examples of software defined power module of the server where fan speed, etc., are controlled based on ambient and server load. However, one can't throw away 1000s of servers overnight to become smarter.
In Machinesense, for Motor driven equipment, we already provide this smart information to the energy auditors. From electrical stand point, a server power feed consists of a motor feed and AC to DC transformer feed. Looking at PF and harmonic signature, it is possible to get energy vs fan speed vs CPU/RAM usage for each server when a single phase Machinesense PA will be used to tap its voltage & current. When this detailed data will be thrown to DCIM, all the relevant data sets will be available for energy saving strategy.
In 2016, Google declared, it has used Deep Mind to reduce server bills by 40%. When thousands of servers will be throwing millions of detailed data, deep learning can help to extract multiple powerful recommendation for energy saving.
Please do write to us, if you have queries (firstname.lastname@example.org).