Article Preview
Top1. Introduction
Elasticity (i.e., the ability to acquire and release resources on demand as a response to changes of application requirements during runtime) is a key feature that drives the popularity of infrastructure clouds (Infrastructure as a Service, or IaaS, clouds). To date, much effort has been dedicated to studying the elasticity of computational resources, which in the context of IaaS clouds is strongly related to the management of virtual machine (VM) instances (Mao & Humphrey, 2011; Marshall, Keahey, & Freeman, 2010; Niu, Zhai, Ma, Tang, & Chen, 2013): when to add and terminate instances, how many and what type to choose, and so forth. Elasticity of storage has gained comparatively little attention, however, despite the fact that applications are becoming increasingly data-intensive and thus need cost-effective means to store and access data.
An important aspect of storage elasticity is the management of I/O access throughput. Traditional IaaS platforms offer little support to address this aspect: users have to manually provision raw virtual disks of predetermined capacity and performance characteristics (i.e., latency and throughput) that can be freely attached to and detached from VM instances (e.g., Amazon Elastic Block Storage (EBS) (AmazonEBS, n.d.)). Naturally, provisioning a slower virtual disk incurs lower costs when compared with using a faster disk; however, this comes at the expense of potentially degraded application performance because of slower I/O operations.
This trade-off has important consequences in the context of large-scale, distributed scientific applications that exhibit an iterative behavior. Such applications often interleave computationally intensive phases with I/O-intensive phases. For example, a majority of high-performance computing (HPC) numerical simulations model the evolution of physical phenomena in time by using a bulk-synchronous approach. This involves a synchronization point at the end of each iteration in order to write intermediate output data about the simulation, as well as periodic checkpoints that are needed for a variety of tasks (Nicolae & Cappello, 2013) such as migration, debugging, and minimizing the amount of lost computation in case of failures. Since many processes share the same storage (e.g., all processes on the same node share the same local disks), this behavior translates to periods of little I/O activity that are interleaved with periods of highly intensive I/O peaks.
Since time to solution is an important concern, users often overprovision faster virtual disks to achieve the best performance during I/O peaks and underuse this expensive throughput outside the I/O peaks. Since scientific applications tend to run in configurations that include a large number of VMs and virtual disks, this waste can quickly get multiplied by scale, prompting the need for an elastic solution.
This paper extends our previous work (Nicolae, Riteau, & Keahey, 2014b) where we introduced an elastic disk throughput solution that can deliver high performance during I/O peaks while minimizing costs related to storage. Our initial proposal focused on the idea of using small, short-lived, and fast virtual disks to temporarily boost the maximum achievable throughput during I/O peaks by acting as a caching layer for larger but slower virtual disks that are used as primary storage. We have shown how this approach can be efficiently achieved in a completely transparent fashion by exposing a specialized block device inside the guest operating system that hides all details of virtual disk management at the lowest level, effectively casting the throughput elasticity as a block-device caching problem where performance is complemented by cost considerations.
In this paper, we complement our previous work by exploring how to predict the performance and cost of running HPC applications that exhibit well-defined I/O behavior. This is a critical challenge, because running such applications at scale requires a massive amount of resources, which makes it important to know in advance whether the desired results can be obtained within a given deadline and cost.
Our contributions can be summarized as follows: