Auto-tuning of Cloud-Based In-Memory Transactional Data Grids via Machine Learning

Abstract

In-memory transactional data grids have revealed extremely suited for cloud based environments, given that they well fit elasticity requirements imposed by the pay-asyou-go cost model. Particularly, the non-reliance on stable storage devices simplifies dynamic resize of these platforms, which typically only involves setting up (or shutting down) some data-cache instance. On the other hand, defining the well suited amount of cache servers to be deployed, and the degree of replication of slices of data, in order to optimize reliability/availability and performance tradeoffs, is far from being a trivial task. As a example, scaling up/down the size of the underlying infrastructure might give rise to scarcely predictable secondary effects on the side of the synchronization protocol adopted to guarantee data consistency while supporting transactional accesses. In this paper we investigate on the usage of machine learning approaches with the aim at providing a means for automatically tuning the data grid configuration, which is achieved via dynamic selection of both the well suited amount of cache servers, and the well suited degree of replication of the data-objects. The final target is to determine configurations that are able to guarantee specific throughput or latency values (such as those established by some SLA), under some specific workload profile/intensity, while minimizing at the same time the cost for the cloud infrastructure. Our proposal has been integrated within an operating environment relying on the well known Infinispan data grid, namely a mainstream open source product by the Red Had JBoss division. Some experimental data are also provided supporting the effectiveness of our proposal, which have been achieved by deploying the data platform on top of Amazon EC2.

Proactive Cloud Management for Highly Heterogeneous Multi-cloud Infrastructures

“…The works in [23], [24] target cloud computing environments using ML-based prediction models to self-tune the runtime configuration of the applications being run on the virtualized infrastructure. Differently from our proposal, these works do not explicitly consider multiple cloud regions.…”

Self-tuning in Distributed Transactional Memory

“…The tuning of the replication degree in a DTM [50,38] is another closely related problem, which encompasses a subtle trade-off between the probability of accessing locally stored data and the cost of the synchronization phase necessary to validate committing transactions. On one hand, in fact, increasing the replication degree generally results into a higher probability that a transaction accesses a data item that is maintained by the local node; on the other hand, for update transactions, it also typically leads to an increase in the number of nodes to be contacted at commit time for validating the transaction and propagating its updates [38].…”

“…Pure black box approaches, instead, have been undertaken in [50,102], where ANN are employed to predict transactions' throughput and response time while varying the number of nodes composing a DTM. In particular, the work in [102] allows for supporting what-if analysis at the granularity of individual transactional classes, and not only on the overall average performance of the entire transactional workload.…”

Optimizing hyperspace hashing via analytical modelling and adaptation

“…Finally, our work is also related to the vast literature in the area of performance modelling of (distributed) database systems, which include a large number of approaches based on queuing theory [31,7,11], and, more recently, on blackbox machine learning methodologies [24,29,12].…”