Extracted Annotations (3/30/2021, 6:15:50 PM)
"We find idle server power demand to be significantly higher than benchmarks from ENERGY STAR and the industry-released SPEC database, and SPEC server configurations—and likely their power scaling—to be atypical of volume servers." (Fuchs et al 2020:379)
"International Data Corporation (IDC) data from December 2014 indicated that volume servers make up the great majority of the US market both for the installed base and annual unit shipments; each year from 2010 to 2015, > 95% of installed servers, by number, were volume servers. Unfortunately, the energy use of volume servers in the field remains poorly understood because they handle heterogeneous computing loads, are quite configurable in terms of hardware and software, and are operated over a wide range of management practices." (Fuchs et al 2020:380)
"We place SPEC and ENERGY STAR servers' energy use in the context of other available data and quantify the difference in power demand between servers in those higher-performing data sets and our best understanding of a typical server operating today" (Fuchs et al 2020:390)
"As such, the power-utilization scaling observed in SPEC and its improvement over time may not actually be representative of server performance under real-world conditions." (Fuchs et al 2020:391)
"ENERGY STAR estimates, which aim to be representative of servers with high volume sales, may not be representative of the actual consumption of these servers once configured after purchase." (Fuchs et al 2020:391)
"Custom-designed volume servers, once an insignificant portion of the server stock in data centers (Brown et al. 2007), are expected to represent about 40% of all volume servers in operating in the USA by 2020, due to the rise of hyperscale data centers managed by large specialized service companies such as Google, Amazon, and Facebook (Shehabi et al. 2016)." (Fuchs et al 2020:391)
"Future work should include field measurements of server power draw in data centers of different sizes, from closets to hyperscale, accounting for variation in configurations that may drive differences in power draw" (Fuchs et al 2020:391)
"servers should include component and setting changes post-purchase, as well as include customized servers used by large internet companies, noting how configurations may vary by the types of computations being performed evaluated" (Fuchs et al 2020:391)
Meta::