CS695 Topics in Virtualizatioan and Cloud Computing Autumn 2014 --------------------------------------------------- Assignment #2 Feedback 1. Title: Workload mix matters (Mix-Aware) Names: Anshu Yadav (124050003), Sanket Arkade (143050089) The G/G/1 mode; in the paper predicts the request rate that can be supported by a given resource allocation (number of servers). in your case, you will need a controller to adjust resources in uni-processor mode. First, you should explore, setup and report the presence of mix of requests. Also, report how is mix-capture with different values of k with clustering. Then, move on to working with mix and SLAs (performance, cost and correctness etc.) 2. Title: Study of memory ballooning and KSM techniques and evaluation of their individual and relative performance. Priyanka Sawant ­ 133059002, Priyanka Naik ­ 144056001 Good description. An interesting set of experiments would be to study impact of ballooning when sharing is ON (shared pages may get ballooned out!) and vice-versa. Can be an input for the ballooning+ksm controller. Scope can get big, focus on simple intuitive controllers and for each building block think of "building-block" experiments. Contact deba, ravali for custom memory-stress program and other workloads. 3. Title: SR-IOV Virtualization Project Kausik and Shubham Title needs to improve. Need to measure both cost and performance benefits in each case. Experiments need detailing. Scope seems limited. - Can add different use-cases to network virtualization options, maybe a SDN use case for the two options and its evaluation, or - Can try bonding driver usage. Two interfaces in guest which are bonded---one throught PV and throught SRIOV and a simple scheduling policy in bonding driver. Can design some evaluation around this as well. 4. Title: Replication-aware Design Aby and John Title needs to be improved. Description is incomplete---no methodology and no design of experiments presented. Providing the interface layer might be ambitious. Try for, - implementing a VMM managed replication service - assume a stateless application in VM and a load balancer that can exploit replicas when created (e.g., a web server). - use VM application or resource statistics for replication decision. (as a first step do this manually) 5. Title: Porting lite­green to KVM Names: Uddhav Arote, Bhaskar Bandyopadhyay - how are you planning to measure power? (contact Swetha, synerg RS) - after host trasnfers VM to server, the host needs to switch to a low-power state (hibernate) and when VM activity increases, host needs to first wake up and then VM can move from server to host. explore magic-packet based wake-on-lan and the use it for above coordination. the above component will need expiremental evaluation. - also, check if you can virtualize an existing host desktop machine, instead of creating new VM for user. 6. Mandar Pawar, Bhavesh Singh - needs title - Lot of state needs to be (accessed) carefully managed for memory similarity estimation. First get a handle on this. - how to get access to all memory pages of a VM? - how to traverse and log its state and compare for similarity? Plan a set of evaluations around this to quantify similarity costs and potential and also to show correctness. Then move on to the controller. Or as fist step, without controller, manually move VMs based on decisions for evaluation. 7. Title: Smart OS - operations and features Team: Pratyush Nalam, Praneeth Naidu - title needs to be improved - description is incomplete and high-level. not clear what will be implemented and evaluation. - first, need to get SmartOS and its components working. this is a major step. - if already not started on this and made progress, will be tricky. - assuming components work, what is the evaluation plan? 8. Topic: Memory management using balloon controllers and live-migration. Team members: Sahil Singh, Chinmay Kulkarni - actual design of experiments not listed. need to think of this carefully. - when to migrate when ballooning can be done? e.g., if balloon is deflated all its memory is provisioned, then will migration help? migration will help only if inspite of balloon based control, memory is still the bottleneck. creating such scenarios for testing and evaluation, needs careful design. 9. Title: Design And Evaluation Of Balloon Controllers Sridhar and Indrajit - what would the balloon controller depend? need experimental evaluation parameters used for decision and their impact. - need a controller that schedules memory over all VMs. e.g., if 3-4 VMs are executing and one of them is stressed, what amount should be gotten from each VM. - need to test the parameters of the controller heuristics as well. 10. Title : Design and evaluation of balloon controllers (in Xen) Sahil and Sagar - what are the parameters on which decisions of the controller will depend on? identify these and formulate 2-3 heuristics for the same. contact Ravali (the TA) for pointers to related work. - need to evaluate impact of heuristic parameters on cost and performance. - need repeatable experiments for comparison across heuristics. 11. Topic: Determine memory and network affinity between VMs for provisioning Deepali and Meenakshi Good description. Need to quantify the impact of affinity (for applications) A generic tool as a deliverable would be good. Takes as input a bunch of VMs and collects their statistics across machines to output affinity values. 12. Pulsar : A guest-driven ballooning implementation for the Xen hypervisor Durgesh Good description. Get to the hypercall implementation quickly. Selecting and giving-away pages from guest will be tricky. Could start with one VM initially and can implement different page selection techniques, ---random, LRU etc.