Reduce Power Consumption in Servers Infrastructure under backup and availability constraints
IT Data centers and global network consume approximately 4 percent power of the overall worldwide power consumption. A global challenging task is to optimize energy consumption in the field of computing, research community is not only focusing on finding ways for reduction of energy cost and maintain profitability, but also working on reduce environment effects of global warming. In this work, we aim to provide power consumption model for servers at data centers with the motive of savings in power utilization allowing servers to go into sleep mode without compromise on Quality of Services (QOS) in terms of availability and performance in terms of delay that occurs in activation from sleep mode when any service arrives at server.
Dr Arshad Ali
Assistant Professor University of Lahore
PhD in Computer Science and Telecommunication
Awais Latif Gouhar
Student University of lahore
University of Lahore MSCS in Computer Science
Power Save in Scenarios
we save power consumption in different Scenarios.
Scenario 1 and 3
Scenario 1 and 3 are similar in terms of power saving, not in configuration. In both two servers are running parallel at a time. We want to put server B (Backup Server) in sleep state when it is at idle state. After up gradation of change in server B we put it in sleep mode for three minutes or more as per requirement and business scenario. Server B remains in idle state most of the time because change is not occurring all the time in server A. We take backup server B from type A server for scenario 1 and 3 as shown in Figure 12.We know that server takes 207 watts when in idle state, takes 17.25 watts in sleep state and server takes 345 watts in full load. In Figure 13red line shows use of real time power consumption of backup server and green line shows power consumption of backup server after applying scenario 1 and 3. In Figure 13, we change transition after 1 hour for understanding and we also observe reduction of power consumption for full day.
How much time server remains active and in sleep mode
It depends on business scenario how much we compromise on time and data loss. But we compromise for 3 minutes. Server A sends backup job to server B after every 3 minutes. Then server B wake ups from sleep mode normally within 25 seconds depending upon server type . Now question arises that for how long server B will remain in active state after wakeup? This depends on DATA and LINK between both servers, i.e., how much data moves from server A to B (change) and nature of link. We analyze SAP servers of company under study. If 100 users are working, then after every 3 minutes, data for backup is about 1 MB which depends on users and work load. Speed of link between server A and server B and network nature (LAN or WAN) determines the time taken for copying data to backup server. We take 1 MB data move in LAN network at worst case within 2 seconds. Today minimum use of cable is cat 5 that can move data in LAN at 10 MB per second. In our scenario 1 or 3, we calculate server B power consumption from sleep to active transition with transfer of 1MB data and then going to sleep mode again. This process takes about 1 minute in worst case. Thus, server takes on average 1 minute to incorporate change after every sleep period of 3 minute. In case, there is no change in server A, then server B does not go to active state and it will remain in sleep state; however we consider worst case scenario to be more realistic. Server B goes in sleep state for maximum of 360 times in 24 hours and thus remains in sleep state for 18 hours and in active state for 6 hours in a day. Total power consumption of 24 hours is power of sleep time + power of active time. Power consumption during sleep mode is equal to 310.5 (18*17.25) watts whereas the same during active period of server is calculated as 2001 (333.5*6) watts. In active period will be synchronized live with backup server and other 18 hours server will remain in idle state. We calculate the power by Pload formula of 50% load that becomes 333.5 Watts. Total power consumption of Backup server B is 5134 Watts for 24 hour periods without applying scenario 1. This is calculated by adding power consumption at 5% load on average and power consumption in idle mode of the server category A by using equation 4.2.The power consumption at average 5% load turns out to be 6.9 watts (2nd part of equation 4.2), thus total power consumption becomes 213.9 watts so 24 hours power consumption is 5134 watts.
In this scenario, there can be more than two servers, five servers, ten or even more. However, we considered 4 servers from category C server’s specifications as mentioned in Table 5in our real setup in order to provide results. If single server load is divided into four servers then load on all servers becomes 25%. Power consumption of single server at 25 % load is 257 watts. Since power in idle mode is 220 watts and 25% load results in more power consumption of 36.75 watts which is 25% of the 147 watts (difference between power at full load and power in idle state). Consequently, power consumption of four servers becomes 1024 watts which is very high keeping in view work load. However, if we apply scenario 2 then only one server is sufficient to handle this data traffic and power consumption of that server at 100 % load will be 367 watts and second server will be at sleep state for high availability which will consume 18.35 watts per hour; thus total power consumption of these two servers becomes 385 watts. Thus, we are in a position to save 639 watts by applying our scenario. In case traffic load is heavy, say 50% or 75%, then we need more servers in active state but we still observe significant power savings. At 50% load, two servers are sufficient to be in active state and one in sleep mode and power consumption after application of our scenario computes to be 752 watts as compared to 1174 watts without this scenario resulting in power saving of 422 watts. Similarly, power saving at 75% load is 201 watts which is still 15% of the total power. On an average, we can save 36% power.
In this scenario we put application server in sleep state when it remains idle for 15 minutes. In small and medium level organizations server are used only during office timings for 8 hour periods and servers remain idle for other 16 hours. Now, we want to put these servers in sleep state. By applying our infrastructure scenario 4 on server’s type A, calculations are similar to those of scenario 1 and 3. Green line in Figure 15 shows power savings. During 24 hours, server remains in sleep state for 16 hours and total power consumption in this state is 276 (17.25 * 16) watts. Power consumption of server during active period of 8 hours turns out to be 2760 watts considering 50% load on server. After applying scenario 4 power consumption of 24 hour is 3036. Total power consumption of category A server is 5192 W without applying any method as already calculated. Thus, we save 2156 watts per server per day.
Our overall and main objective is to devise a mechanism focusing on reduction of energy consumption in server infrastructure and decreasing environmental effects. We want to
Reduce Servers Power
In Data Centers
Propose a scheme to reduce power consumption in DR, backup and application servers.
by sleep mode
Enable sleep mode of DR, backup and application servers during their idle time which may have a small compromise on backup and availability.
Power consumption is main factor in current and future server infrastructure. We pointed out four different problems where power is wasted after observing four scenarios being used in every small and large scale enterprise server’s infrastructures. In scenario 1, we noticed 55 % power saving after applying sleep mode for three minutes in additional backup server. Thus, power saving of 65% is observed with compromise on availability of backup server for maximum 3 minutes. In scenario 2, servers are used for cloud computing or application services. In one set there can be 2 to hundreds of servers that can provide cloud services. We allowed servers to go in sleep mode and shutdown mode as per requirement. Similarly, scenario 2, we found 35.8% savings of power consumption without compromising on quality of services and SLA. In scenario 3which is similar to scenario 1, again we save 55% power in backup server by changing the methodology and configurations of active servers. In our last scenario, our focus was on those organizations which use their data servers for 9 hours per day and servers remain in idle state for rest of the time and we observed 41.5% savings in this scenario too. .In summary, there exists huge potential of power savings in server infrastructure of data centers and the same can be exploited by allowing servers to go in sleep mode during their idle periods.
 Gladisch, Andreas, Christoph Lange, and Ralph Leppla. "Power efficiency of optical versus electronic access networks." Proc. of ECOC, Brussels, Belgium (2008).Andreas - Belgium (2008)
 Kelly, Tim, and Standardization Head. "ICTs and climate change." ITU-T Technology, Tech. Rep (2007).Tim - ITU-T Technology
 Wang, Shengquan, et al. “Powersleep: a smart powersaving scheme with sleep for servers under response time constraint.” Emerging and Selected Topics in Circuits and Systems, IEEE Journal on 1.3 (2011): 289-298.Shengquan - IEEE Journal on 1.3 (2011)
 Gandhi, Anshul, MorHarchol-Balter, and Michael A. Kozuch. "Are sleep states effective in data centers?." Green Computing Conference (IGCC), 2012 International. IEEE, 2012.Anshul - IEEE, 2012
select scenario for calculations of power consumption