wiki:public/20122013

Project: Massive Parallel Computing For The Robust Tenant Placement

Team: Tim Januschowski (SAP Innovation Center), Jan Schaffner (Hasso Plattner Institut)

Research institution: SAP Innovation Center Potsdam

Abstract: In the low-margin Database-as-a-Service (DBaaS) industry, most ser- vice provider leverage economies of scale by consolidating multiple customers (tenants) on a single server. The Robust Tenant Placement Problem (RTP) is one key challenge that needs to be addressed when aiming for a cost- effective operation of a DBaaS server farm in a multi-tenancy scenario. The task in the RTP is to place replicated tenants on as few servers as possi- ble while maintaining Service Level Objectives e.g., on response times and availability.
In previous work, we considered heuristics for solving the NP-complete RTP. Currently, we only have a limited understanding of the gap between our heuristically found solutions and the optimal solution. Our approach is to obtain lower bounds on the optimal solutions from a Mixed Integer Programming (MIP) formulation for the non-linear RTP. However, due to the complexity of the RTP, we have not been able to tackle instances with practically relevant size with MIP solvers like CPLEX. In our project, we propose to use the 1000 core server to (hopefully) solve some RTP instances of relevant size. Contrary to traditional clusters where achieving a high degree of parallelization for a MIP solver involves deep engineering, the deployment of a 1000 core machine would allow us to rely on standard tools, such as CPLEX which allows a high degree of parallelization out of the box.
Using the entire set of CPUs for up to two months would allow us to solve hopefully at least one relevant instance of the RTP. A solution of one or more practically relevant instances, enabled by the massive parallel computing power of the 1000 core server, would lead to a much deeper analysis for our practically viable heuristics of the RTP and the problem itself.

Last modified 6 years ago Last modified on Nov 20, 2012 4:44:23 PM