服务器能源效率探究：2016年的现状与挑战
178 浏览量
更新于2024-08-28
收藏 674KB PDF 举报
"Energy Proportional Servers: Where Are We in 2016?"
在2016年，数据中心的能源消耗问题已经变得至关重要。由于巨大的能源消耗，不仅导致电费高昂，还产生了大量的碳排放，对环境造成了显著影响。尽管当前的服务器和顶级互联网公司的数据中心在能效方面取得了前所未有的进步，但外部工作负载的波动和内部资源利用率的变化仍然提出了对能源比例计算的需求。了解服务器的能源比例性有助于优化工作负载分配，同时减少能源消耗。
在论文《Energy Proportional Servers: Where Are We in 2016?》中，作者Congfeng Jiang、Yumei Wang、Dongyang Ou（来自杭州电子科技大学）、Bing Luo和Weisong Shi（来自韦恩州立大学）深入研究了这一主题。他们分析了2007年至2016年第三季度所有477个有效的SPECpower_ssj基准测试结果，并根据硬件可用年份重新组织这些数据，以更准确地分析生产服务器的性能。
通过全面分析，研究者发现以下关键点：
1) 尽管存在所谓的能源比例性的停滞现象，但服务器的实际能效仍有所提高，这表明在硬件设计和优化方面取得了一定的进步。
2) 随着技术的发展，服务器在保持高性能的同时，其能源效率也在持续提升，这归功于新的架构和组件的设计。
3) 工作负载的动态变化对服务器能源效率的影响显著，需要智能的管理和调度策略来平衡性能与能耗。
4) 服务器的能源比例性不仅受硬件影响，软件层面的优化同样重要，例如操作系统和应用程序的能源效率改进也是降低能耗的关键因素。
5) 数据中心的冷却系统和整体设计对能源消耗也有显著影响，优化这些方面可以进一步提高能效。
研究还指出，为了实现真正的能源比例计算，需要在硬件设计、软件优化以及工作负载管理等方面进行综合考虑。这包括但不限于采用更高效的处理器、改进的内存和存储系统、智能电源管理系统，以及利用虚拟化技术进行资源调度。此外，对服务器的全生命周期能源效率进行评估也至关重要，包括制造、运行和废弃阶段的能耗。
这篇论文揭示了尽管数据中心的能效有所提升，但仍存在挑战，尤其是在面对工作负载波动时如何维持能源比例性。研究者的工作为未来的研究和实践提供了有价值的洞察，以推动服务器和数据中心向更加节能的方向发展。
a SPECpower result, the energy efficiency values are entitled as
performance to power ratio.
(4) Server overall energy efficiency. Server overall energy
efficiency is overall performance to power ratio of a server in a
SPECpower result, i.e., the ratio of sum of ssj_ops over sum of
power for 10 utilization levels (from 10% to 100%) and active
idle. The server overall energy efficiency is also referred as a
server’s SPECpower score.
(5) Peak energy efficiency. Peak energy efficiency is
defined as the greatest (peak) energy efficiency of a server
among all utilization levels.
(6) Energy proportionality (EP). In this paper we use
the energy proportionality metric in [14]. Take a server from
SPECpower dataset in 2016 with overall score 12212 as an
example. We plot its power-utilization curve in Fig.1. Note that
the power in Fig.1 is normalized to its power at 100%
utilization. The solid line is the energy proportionality curve of
the sample server and the dotted line is of an ideally energy
proportional server. With the power-utilization curve in Fig.1,
we can compute the energy proportionality of the server as
following [14]:
ൌ ͳ െ
ೝೌ
ି
ೌ
ೌ
(1)
Thus the power-utilization curve in Fig.1 is also called the
energy proportionality curve. From Eq.1 we can see that EP is a
value equal to or greater than zero but less than 2.0. For an
ideally energy proportional server, its EP value is 1.0. For the
server in Fig.1, we approximate its EP by summarizing the
areas of ten trapezoids corresponding to ten utilization intervals
and then get its EP value 1.02 according to Eq.1
III. E
VOLUTIONARY TREND ON ENERGY PROPORTIONALITY
A.
Energy Proportionality Evolution
During the last decade, server energy efficiency improves
significantly thanks to technical breakthroughs including low
power electronics and designs, dynamic voltage frequency
scaling, and cross level coordinated power savings. We
compute the energy proportionality value of each server
according to Eq.1 and present the trend curve of all the servers’
EP and EE values in Fig.2 and their statistics in Fig.3 and Fig.4.
Not surprisingly, both energy proportionality and energy
efficiency improve with times. The average energy
proportionality increased significantly from 0.30 to 0.82 from
2005 to 2012 and seems stagnant at 0.84 in 2016. Similarly, the
minimal EP in each year also improves significantly. For
example, newest servers made in 2016 have minimal EP of
0.73, which is the greatest EP value in 2009. Among 477
servers, the least energy proportionality value is 0.18 (in 2008)
and the highest energy proportionality value is 1.05 (in 2012).
From Fig.2 and Fig.3 we observe that there are two significant
improvement steps in EP, one is from 2008 to 2009, and
another is from 2011 to 2012. From 2008 to 2009, the average
and median EP increase from 0.37 to 0.55 (+48.65%), and 0.37
to 0.56 (+51.35%), respectively.
Similarly, from 2011 to 2012, the average and median EP
increase from 0.66 to 0.82 (+24.24%), and 0.67 to 0.85
(+26.87%), respectively. From 2008 to 2009, the majority of
the servers switch their processor microarchitecture from Core
(Penryn) to Nehalem. From 2011 to 2012, the majority of the
servers switch their processor microarchitecture from Nehalem
(Westmere) to Sandy Bridge. These two switches are called
tock in Intel’s tick-tock chip iteration model which designates a
new microarchitecture.
Although energy proportionality fluctuates with times in
Fig.3, their energy efficiency increases monotonically with
times in Fig.4, including its average value, maximal value and
median value of each year in all three metrics. Only the
minimal values of the three metrics in 2014 decrease because
one server in 2014 has low EE as 1469 and low EP as 0.32.
This server has a tower form factor and Intel Core i5-4570
processor usually not designed for server purpose use, which is
quite different from other servers in 2014 or other years. Since
there are few results for year 2004, 2005, 2006, and 2014, the
statistical results in these years pose outliners and greater
deviations. For example, EP decreases significantly from 2004
to 2005 and from 2012 to 2014. But the median EP in 2014 still
increases compared with 2013.
We plot the CDF chart of energy proportionality in Fig. 5.
Among 477 servers, 25.21% servers have EP between 0.6 and
0.7, and 17.44% servers have EP between 0.8 and 0.9. These
two intervals have the most EP values. Moreover, 99.58%
servers have EP less than 1.0.
Fig. 1. Energy proportionality curve.
Fig.2. EP and EE evolution.
Fig.3. Stats. trend of EP.
0
0.5
1
0% 20% 40% 60% 80% 100%
Power
Utilization
2016 EP=1.02
Ideal
0.0
0.2
0.4
0.6
0.8
1.0
1.2
0
2000
4000
6000
8000
10000
12000
14000
2004 2010 2012 2015
EP
EE
EE
EP
0.0
0.2
0.4
0.6
0.8
1.0
1.2
2004 2007 2010 2013 201
6
EP
Max
Median
Average
Min
251925161651
剩余11页未读，继续阅读
2013-11-26 上传
2021-11-22 上传
2021-05-29 上传
2021-05-14 上传
2021-05-01 上传
2021-05-12 上传
2020-02-02 上传
2020-02-19 上传
2019-08-16 上传
weixin_38602098
- 粉丝: 3
- 资源: 963
我的内容管理
展开
