Enabling Resource Sharing between Transactional and Batch Workloads Using Dynamic Application Placement

Abstract

Abstract. We present a technique that enables existing middleware to fairly manage mixed workloads: batch jobs and transactional applications. The technique leverages a generic application placement controller, which dynamically allocates compute resources to application instances. The controller works towards a fairness goal while also trying to maximize individual workload performance. We use relative performance functions to drive the application placement controller. Such functions are derived from workload-specific performance models-in the case of transactional workloads, we use queuing theory to build the performance model. For batch workloads, we evaluate a candidate placement by calculating long-term estimates of the completion times that are achievable with that placement according to a scheduling policy. In this paper, we propose a lowest relative performance first scheduling policy as a way to also achieve fair resource allocation among batch jobs. Our technique permits collocation of the workload types on the same physical hardware, and leverages control mechanisms such as suspension and migration to perform online system reconfiguration. In our experiments we demonstrate that our technique maximizes mixed workload performance while providing service differentiation based on highlevel performance goals.

Resource pool management: Reactive versus proactive or let’s be friends

“…Commonly available resource demand traces are the basis for our management system. However, there are many research papers, e. g., [36,37], which design dynamic provisioning systems for targeted classes of applications, e. g., multi-tier applications. There are excellent earlier works on load sharing systems that support batch-like workloads [38,39].…”

Fastrack for taming burstiness and saving power in multi-tiered systems

“…The mathematical definition of I in Eq. (1) shows that the index of dispersion jointly captures variability and burstiness in a single number and that the summation of autocorrelations at all lags further gives a measure of the strength of burstiness.…”

“…Commonly available resource demand traces are the basis for such management system. However, there are many research papers, e. g., [11,13,1], which design dynamic provisioning systems for targeted classes of applications, e. g., multi-tier applications. In [11], the authors discuss system-workload context (system and workload characteristics) that impact energy saving policies.…”

To migrate or to wait: Bandwidth-latency tradeoff in opportunistic scheduling of parallel tasks

“…Existing work can be divided into three categories: (i) complete control, where both types of jobs are scheduled jointly [4], (ii) one-sided control with complete knowledge, where the scheduler of low-priority jobs faces fluctuating resources but has full knowledge of the fluctuations throughout time [5], and (iii) one-sided control with partial knowledge, as in this paper, where the knowledge of resource fluctuations is limited to past observations and long-term statistics. The last case is most generally applicable due to its non-intrusiveness and minimalistic assumption, but it also suffers the most performance loss.…”

“…Given the reward Ri(r) for Ti as defined in (4), it is easy to see that the transition between policies must occur at the intersecting points between Ri (r) 's. Since Ri(r) = Pi -'YMi, the intersecting point between Ti -l and Ti is given by 'Yi = (Pi -Pi-l)/(Mi -Mi-d…”