This year was our first attendance at IEEE Cloud. We did not really know what to expect. This is the first of 2-3 posts which capture some of what went on at the conference.
IEEE Cloud is one conference which is is part of a larger suite of conferences covering diverse aspects of computing services (including the International Conference on Web Services, International Conference on Mobile Services and the International Congress on Big Data). It was structured such that there were single tracks dedicated to each individual conference but the keynotes and the panels spanned all the conferences. There were a few hundred folks in attendance, mostly coming from academia. We missed a few days of the conference but we’ll spend enough time here to get to know some of the folks and get a feel for it (we can’t afford the time to spend 6 days at a conference!).
On day 4 – June 30 – we spent most of our time at the Applications and Industry Track. We attended sessions on Map-Reduce, Cloud Scalability and Clouds for High Performance Computing. There was also a keynote by Munindar Singh on developing a science of security as well as a panel session which tried to answer questions relating to the next seven years in ‘Services Computing’.
While Singh’s keynote was entertaining, its security focus was a bit of a distance from what we’re doing in ICCLab. The panelists mostly took their perspective for future trends from the academic publishing world, pointing out both Big Data and IoT as the main drivers of services in the future.
Of the sessions we attended, we found the talk by Thomas Goldschmidt of ABB Corporate Research on handling large amounts of energy data generated in an industrial context. There, he noted that KairosDB on Cassandra offered a good solution for scalable, reliable, high performance storage and subsequent data access for their quite demanding data requirements. A talk by Zhenyun Zhuang of LinkedIn on performance issues in the JVM was also interesting; there, it was noted that Transparent Huge Pages which are a relatively recent introduction to linux can have a detrimental impact on JVM performance as the JVM is not aware of this mechanism – the authors proposed a mechanism to determine when to activate and deactivate THP to improve JVM performance. Finally, Andrew Younge of Indiana University gave an interesting talk on GPU performance in virtualized environments. In his talk, Andrew presented results which demonstrate that on modern hardware, widely used hypervisors (VMWare, Xen, KVM) deliver performance which is almost the same as performance obtained when running applications on bare metal – indeed VMWare was shown to deliver better performance than running the application on bare metal. This gave rise to the notion that GPU supports in COTS hardware and hypervisors are good enough to support a significant amount of HPC applications in clouds.
Overall, quite an interesting day…more news tomorrow.