Paper: Energy-aware server provisioning and load dispatching for connection-intensive internet services
Summary + Liked:
• The idea of using a hybrid approach of load balancing and load skewing would lead to better results as we can see the load patterns are sinusoidal. Hence in inclination time frame, load balancing approach would be useful in distributing load while turning on new servers. In decline time frame load skewing would be preferred such that SID would be less and servers can be turned off.
• Selecting turnoff server on basis of current number of connection hosted by that server help in loosing minimum connection at maximum energy savings. Also approach of rotation in selecting which server to turn off can ensure reliability of hardware.
• Paper presents a trade-offs between Energy savings and Quality of service. Also it analyzes various parameter which plays significant role in determining connection load. It uses login rate, number of connection, upper bound of these parameters and some heuristic parameters to dynamically determine server-provisioning value.
• Preference to hysteresis based provisioning over forecast based approach in case of lack of data is justified by better results. Paper also presents comparison between skewing vs. load balancing approach and forecasting vs. hysteresis approach.
• Results shows that skewing and forecasting based approach leads to 30% approx energy savings.
Disliked:
• Paper presents a loose bound (not a tight bound) on server provisioning. Also interdependence of various tuning parameter is not clearly specified.
• Why can’t server do migration of all the connections using live migration approaches in ongoing research?
• Though paper presents a new approach for short term forecasting but this model assumes that values in functions are highly correlated. How this model takes care of burst scenario is not clear?
• Approach, heuristic and data presented in paper are specific to Windows live only.
details will be provided later.. right now i dont have any time
No comments:
Post a Comment