A Performance Model of Multi-Version Concurrency Control

Abstract

In this article we present a performance model for MultiVersion Concurrency Control (MVCC)

Analytical modeling of lock-based concurrency control with arbitrary transaction data access patterns

“…Some of the previous results differ from our proposal in that they cope with modeling of optimistic concurrency control, while our interest is in 2PL. Similar considerations apply to the work in [13], where a performance model for snapshot-isolation Multi-Version concurrency control is provided, which is intrinsically different from 2PL and provides a different, non-serializable transaction isolation level [8]. On the other hand, compared to literature analytical results related to lock-based concurrency control (see, e.g., [20,21,23,26]), our proposal exhibits the innovative features of capturing the effects of data access locality vs specific transaction execution phases, and removes the assumption of lock access queue bounded by one unit.…”

“…For this concurrency control protocol we present in this article an innovative analytical model, which is able to overcome several limitations characterizing any preexisting analytical result we are aware of (e.g. [6,13,15,23,26]), thus providing a more accurate tool for the performance analysis of 2PL based systems.…”

Performance Modelling of Partially Replicated In-Memory Transactional Stores

“…Nevertheless, these works typically rely on assumptions on the workload or on simplifications that can hinder their accuracy in real environments. These include: uniform data access patterns [14,15] or detailed apriori knowledge of the distribution of access to data [8,16]- [18]; extreme replication policies, i.e., full [19,20] or no replication [21]; scale-independent probabilities of accessing remote data [22] -which do not match the dynamics induced by popular distribution policies like consistent hashing; constant communication delays [9,19] or detailed knowledge of the underlying network topology/protocols [13,23,24].…”

“…In order to obtain the expected number of aborts for an update transactions, we need to compute the lock conflict probability. To this end, like in previous works (e.g., [14,15,20]), each lock is modeled as an independent M/G/1 server: a lock is acquired at a rate λ lock and the average service time to complete a request, T h , is the lock hold time, namely the time since the lock acquisition and until its release. Note that, in general, a transaction can hold more than one lock at a single time and, as we shall see, hold times for locks taken by the same transaction are correlated: such characteristics violate the independence assumption for the aforementioned server.…”

Identifying the Optimal Level of Parallelism in Transactional Memory Applications

“…In the area of performance modelling of STM, existing literature can be subdivided in solutions based either on analytical techniques [11,12], or on statistical methods [13,14]. Solutions based on analytical models have a good extrapolation power, namely they typically exhibit good accuracy even in workload/scale scenarios not previously explored.…”