Experiences and Additional Research Sample Clauses

Experiences and Additional Research. We have presented preliminary experiences and observations based on our exercise to deploy two production CFD codes on four different target platforms characterized by heterogeneity in secondary attributes. Noteworthy is the effort required for preparing the target platforms for the execution including provisioning of the application, required packages and libraries installation and other logistical hurdles. In this study, we softconditioned all machines manually and installed only the necessary and sufficient packages. We observed that provisioning of a machine took about a day for an experienced member of the development team; in addition, multiple requests and interactions with system administrators were needed. It significantly hampers switching between machines. ADAPT project aims at unsupervised software deployment and in Chapter 6 we show how this mundane and unproductive task can be handle nimbly and automatically. Comparing on-premise and on-demand targets for the applications we tested, we found some evidence to support the claim that IaaS resources may be utilized for scientific CFD simulations possibly at lower cost than incurred locally. In particular, the spot-request feature coupled with availability of cutting edge resources (16-core nodes, 60GB RAM as opposed to 3-year old, 2–4 core nodes with 4GB RAM), suggests that small on-demand assemblies may be a viable alternative to local clusters. It is not without significance that IaaS’s provide resources immediately, while local and grid resources are often subject to long queue wait times—an aspect that might offset any additional expense. Another factor is size; at least in our case, only Cloud providers could provide a large enough offering to sustain the biggest, 1000-core task. Furthermore, while a modern local computing cluster, with an efficient interconnection network will outperform an on-demand assembly (which is highly vulnerable to network performance), the cloud solution might be useful for other reasons. The cost and performance considerations open other research areas. The first extension should explore possible performance tuning. The SaE applications are by-design developed for HPC platforms and this assumption does not have to work well on other platforms such as clouds; investigation related to performance adaptation is presented in Chapter 4. Another research topic is inspired by utility computing ideas: if an application can be executed on various targets that have differen...