|
|
|
|
|
I/O Characteristics and Implications of Big Data Processing on Virtualized Environments |
|
PP: 591-598 |
|
Author(s) |
|
Sewoog Kim,
Dongwoo Kang,
Jongmoo Choi,
|
|
Abstract |
|
In this paper, I/O characteristics of data-intensive applications running on virtualized environments are explored. It is
observed that virtual machines have a tendency to request I/Os in a bursty manner. Also, I/Os are triggered by several virtual machines
at the same time. These concurrent and bursty I/Os cause the interference among virtual machines such as frequent context switches and
long seek distances, which eventually deteriorates I/O performance significantly. To overcome this problem, a novel burstiness-aware
I/O scheduler is proposed, which consists of three components: burstiness detector, coarse-grained dispatcher and starvation handler.
The key idea of the proposed scheduler is detecting bursty virtual machines on-line and allowing a detected machine to consume most
of disk bandwidth exclusively for a given time quantum to reduce the interference. In addition, it provides long-term fairness while
avoiding starvation. Performance evaluation based on implementation shows that the proposed scheduler can improve the execution
time and I/O throughput of the three real workloads by decreasing the number of context switches and seek distances. |
|
|
|
|
|