database systems on virtual machines: how much do we lose?
DESCRIPTION
Database Systems on Virtual Machines: How Much Do We Lose?. Kristin Travis March 2, 2011. Article Information. Authors: Minhas, Yadav, Aboulnaga, Salem University of Waterloo Conference: IEEE 24 th International Conference on Data Engineering Date: April 2008. Background. - PowerPoint PPT PresentationTRANSCRIPT
Database Systems on Virtual Machines: How
Much Do We Lose?
Kristin Travis
March 2, 2011
Article Information
Authors: Minhas, Yadav, Aboulnaga, Salem University of Waterloo
Conference: IEEE 24th International Conference on Data Engineering
Date: April 2008
Background
Obvious benefits with using virtual machines Decreased costs, flexibility, etc.
Virtual machines add performance overhead Question: is it worth it? Personal interest
Setup
DBMS: PostgreSQL 8.1.3 Base System: SUSE Linux 10.1 VM system uses VMM Xen 3.1
VM also runs SUSE Linux 3 GB memory, 1 virtual CPU Test DB is 2 GB in size, ensuring that the entire DB
can fit in memory
22 TPC-H queries are run in identical settings on both machines
First Experiment “Warm” Experiment
Run the 22 queries to “warm” up the buffer, then measure performance in the 2nd run
Queries experience a fairly large overhead Slowdown in system time is much larger than slowdown in user time
Base
Runtime
(secs)
Xen
Runtime
(secs)
Relative
Slow-down (%)
Q3 4.61 5.82 26.23
Q9 10.52 11.36 7.91
Q13 13.36 14.1 5.59
Q18 8.86 10.13 14.36
Q21 2.3 2.48 7.84
Overhead Causes
System Call Time System calls are usually handled by OS, but in the VM
most go through the Xen VMM Negligible in this case
Page Fault Handling Time Page faults in Xen are more than twice as expensive
as they are in the base system
The average relative slowdown of all 22 queries is only 9.8%
Second Experiment “Cold” Experiment
Restart PostgreSQL and flush the file system caches before running each query
Interesting results: the relative slowdown is not high, and some queries actually run faster
Slowdown is caused by disk I/O, having to read from the disk every query
Some queries run faster because Xen VMM aggressively prefetches data for the VM
The average relative slowdown of all 22 queries is only 6.2% (better than “warm” experiment)
Conclusions
The many advantages to running a DBMS in a VM does not come at a high cost in performance
Future work: Different VM environments (Hyper-V) Different DBMS (MySQL, Oracle, etc.) Using multiple VMs on one physical machine