06-10-2016, 11:47 AM
1458010328-CapNetcritique.docx (Size: 16.5 KB / Downloads: 5)
The drive of the authors of this paper was to develop a real-time wireless network for power capping, an emergency and time-critical event by replacing the existing wired networking solutions in Data center management systems (DCM). The complications of using wired networking solutions have been noticeably described in the paper. The large scale of operations in DCMs, high cost and the prominence of avoiding human errors lead the authors to think beyond and upgrade the prevailing cabled networks using the low power and low cost wireless technology.
The background of how power capping is done in the wired networking and the problems occurring during this process have been clearly explained in the paper which gave the authors enough motivation to implement the use of wireless networks in power capping. The idea to use wireless controllers and receivers by replacing the wired DiGi switches and wired controllers reduced the costs expressively and the cost model mentioned in the paper explaining the cost difference between wired DCM and CapNet is noteworthy.
The approach to use low bandwidth wireless like ZigBee (IEEE 802.15.4) instead of IEEE 802.11 and the validation given for it is highly convincing. The implementation of 802.15.4 simplifies the communication protocols and also reduces the data rate when compared to 802.11. The in-depth study performed at the data center of Microsoft corporation to check the radio environment of the wireless DCM inside the racks is one of the major strengths of the research paper. The results derived by the authors of this paper with the help of the above mentioned IT giant is based on real-time data rather and is not based on assumptions. The effort the authors have managed to put into this research should be respected.
The step by step process of the working of CapNet has been clearly explained in the paper. Firstly, the use of sliding window and collision avoidance approaches to gather power measurements from the servers and next, issuing the power capping commands to subset of the servers to maintain the aggregate power consumption of the cluster less than the cap of the cluster. By using these approaches, the research has proven that CapNet can meet the real-time requirements of power capping.
The distributed event detection policy used by the authors to develop a new power capping protocol to capture the dynamic changes and predict the upcoming power peak in the servers avoiding the use of historic readings and predefined communication schedule has achieved the expected results for the authors.
The experimental results in the research paper are keenly observed and noted appropriately for every metric and these results would be supportive for the further research in this proposed wireless mechanism.
The few limitations which I feel are in the current research and needs to be looked into future work is that theexperiment needs to be conducted for longer time rather than six months so that the data traces which were used would have been more information revealing. Another limitation what I feel is during the process of power capping, in the aggregation and control phases, the process of handling the broadcast failures and handling false alarms can be improved so that power capping is done much faster before reaching the trip time of the circuit breaker (CB). And also the number of servers used in this research is decent but the evaluation can be more enthralling, when the count of servers is higher.