The Environmental Protection Agency (EPA) estimates that energy consumption at data centers could grow to 100 billion KWhr, contributing to 2.9% of the total US electricity needs, by the year 2011. In view of the growing energy requirements of data centers that directly contribute to the Total Cost of Ownership (TCO), Barroso and Holzle have made the case for energy proportionality within future computing systems. While this goal is relatively easier to achieve for subsystems with components that are energy proportional (e.g., CPU), it is substantially more difficult with inherently non energy-proportional disk devices. Furthermore, applying classical server consolidation techniques such as virtual machine (VM) migration to storage as-is is not practical because it would involve large amounts of data movement for similar migration operations.
The goal of this project is to build energy-proportional storage systems that consume energy in proportion to the I/O workload intensity. The high-level approach we propose is dynamic workload consolidation on a subset of the physical storage devices combined with transitioning the remaining devices to low power states in response to changes in workload intensity.
Below, we present a set of multi-day block traces taken on real running system. All the systems were running Linux using the ext3 file system.
|home1.tar.gz home2.tar.gz home3.tar.gz home4.tar.gz||Four different end-user / developer home directories.|
|mail.tar.gz||Mail server of our department using Postfix. Only the INBOX accesses are recorded.|
|online.tar.gz||Course management system of our department using Moodle.|
|webmail.tar.gz||Web interface to the department mail server.|
|webresearch.tar.gz||Web-based management of around 10 FIU research projects using Apache web server.|
|webusers.tar.gz||Web server hosting faculty, staff, and graduate student web sites.|
The traces files (one per day) are in ASCII and each record is as follows:
[ts in ns] [pid] [process] [lba] [size in 512 Bytes blocks] [Write or Read] [major device number] [minor device number] [MD5 per 4096 Bytes]
In the case of the home traces the format is different for the digests:
[ts in ns] [pid] [process] [lba] [size in 512 Bytes blocks] [Write or Read] [major device number] [minor device number] [MD5 per 512 Bytes]
This work is supported by the National Science Foundation under grant CNS-1018262 and by IBM Corporation.