Distributed Servers do spend a lot of time doing nothing.

Some of you may recall the debate I had with a Sun engineer  here over the merits of the mainframe v distributed servers.   One of his points was that while his arguement was bolstered by "standard" benchmarks which IBM does not publish for the mainframe our consolidation claims were based on servers that were largely idle. 

Well, in a recent survey covering  about 3000  Intel servers of various types,  the peak to average  utilization ratios were found to be  between 6 and 7.5.   This means the average utilization on those servers is between 13 and 17% assuming they all peak at 100%.  Given that peaks are often managed to be lower than 100% for any sustained interval, the utilization could well be lower on average than the numbers indicate.   .

Also we have looked at our zVM clients and found that they run between 4 and 50 VMs per IFL. With the 2 sigma box around the mean spanning from 4 to about 22.   The finest grained benchmark we have in hand that does virtualization is for 12 images per IFL.   Most benchmarks are run withmultiple cores per OS image and "on the metal" rather than virtualized.  They also are driven with very low workload variability  (P/A approaching 1).   This is why they are completely invalid for representing consolidated and virtualized work.

The bottom line is that distributed servers do spend a lot of time doing nothing.  Well, they do generate some heat... 

by Joe Temple March 10, 2010
Permalink | Comments (5) | TrackBack (0)



The postings on this site are our own and don’t necessarily represent the positions, strategies or opinions of our employers.
© Copyright 2005 the respective authors of the Mainframe Weblog.