<html><head><meta http-equiv="Content-Type" content="text/html; charset=utf-8"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;" class=""><div dir="auto" style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;" class=""><div class=""><br class=""></div><b class="">This is a reminder of Chaojie (Sam) Zhang's MS Presentation.</b><br class=""><br class="">------------------------------------------------------------------------------<br class="">Date: Monday, August 6, 2018<br class=""><br class="">Time: 3:00 PM<br class=""><br class="">Place: Ryerson 255<br class=""><br class="">M.S. Candidate: Chaojie Zhang<br class=""><br class="">M.S. Paper Title: MANAGING THE VALUE OF VOLATILE CLOUD RESOURCES:<br class="">INFORMATION DISCLOSURE AND GUARANTEE-PRESERVING MANAGEMENT<br class=""><br class="">Abstract:<br class="">Cloud providers sell unreliable or �volatile� resources that are<br class="">unused by foreground (reserved/high priority) workloads. The value<br class="">users can extract from these resources depends on the (i) volatile<br class="">resource management algorithm, and (ii) information provided to users<br class="">about the resources as statistical guarantees. We describe and<br class="">evaluate four volatile resource management approaches (Random, FIFO,<br class="">LIFO, LIFO-pools) using commercial cloud resource traces from 608<br class="">Amazon EC2 instance pools, for a 3 month period from 5/2017 to 8/2017<br class="">from Amazon�s four US regions, each of which contains 2-6 availability<br class="">zones. We also consider the value of several information models (MTTR,<br class="">limited statistics, Full distribution, and Oracle) that statistically<br class="">characterize the resources. Our results show volatile resource<br class="">management algorithms can increase user value by 30 to 45% in four<br class="">instance exemplars. For example, LIFO and FIFO vs Random can make more<br class="">than 2x difference. Slightly richer information models (90pctile)<br class="">combined with LIFO and LIFO-pools volatile resource management<br class="">increase user value by as much as 10-fold. Our results suggest that<br class="">cloud providers should pay significant attention to what statistical<br class="">information they provide to users. Then our results of relative user<br class="">value per resource-hour for the four exemplars show that 90pctile<br class="">information model is best in all cases and achieves close to the max<br class="">possible. Hence, simple statstics , such as 90th percentile, can<br class="">increase achievable value by 10% to as much as 5x. However, skewed<br class="">distribution can lead to misleading information, and thus sharply<br class="">reducing derived value. And, these results broadly characterize the<br class="">vast majority (475 of 608) of instance pools. The results are the same<br class="">ordering for VRMs and information models as in two exemplars, and the<br class="">frequency of various relations that are key conclusions for the<br class="">exemplars are carried along in the vast majority of instance pools.<br class="">Futhermore, we provide a detailed drill-down showing how the volatile<br class="">resource management algorithms affect resource interval durations, and<br class="">thus potential user value. We further show how the information model<br class="">shapes user targeting, success rate, and user value. Finally, we study<br class="">offline and two online algorithms to maintain statistical guarantees<br class="">in the face of foreground load changes. Our offline algorithm results<br class="">show that it is feasible to fully preserve statistical guarantees<br class="">under foreground load changes by delaying the release of each volatile<br class="">resources to users by a short period of time with trivial resource<br class="">waste, from 0.5% to 8.9% in extreme cases, at the same time increasing<br class="">user value by up to 134%. Moreover, two online algorithms, AIMD<br class="">algorithm and Distribution Targeting algorithm, can dynamically<br class="">preserve guarantees with modest resource waste, and in doing so<br class="">increase user value by up to 82%. This suggests that further research<br class="">exploring online algorithms is promising.<br class=""><br class="">Chaojie's advisor is Prof. Andrew Chien<br class=""><br class="">Login to the Computer Science Department website for details:<br class=""><a href="https://www.cs.uchicago.edu/phd/ms_announcements#chaojie" class="">https://www.cs.uchicago.edu/phd/ms_announcements#chaojie</a><br class=""><br class="">=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=<br class="">Margaret P. Jaffey margaret at <a href="http://cs.uchicago.edu/" class="">cs.uchicago.edu</a><br class="">Department of Computer Science<br class="">Student Support Rep (Ry 156) (773) 702-6011<br class="">The University of Chicago <a href="http://www.cs.uchicago.edu/" class="">http://www.cs.uchicago.edu</a><br class="">=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=</div></body></html>