Browse Prior Art Database

Optimal IO path checker

IP.com Disclosure Number: IPCOM000200261D
Publication Date: 2010-Oct-03
Document File: 2 page(s) / 20K

Publishing Venue

The IP.com Prior Art Database

Abstract

The common practice to measure the maximal IOPS to a storage controller, is to generate a stress test and measure the overall IOPS over time. This article suggest an alternative method, in which the storage system returns immediately an acknoledge for an IO request minimizing the system contribution to the latency, in order to detect the maximal IOPS possible under the given environment which is beyond the scope of the storage controller (host hardware and configuration, network equipment and bandwidth (e.g. SAN)

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 51% of the total text.

Page 01 of 2

We encounter cases where it is required to check for the
maximum IOPS number available between a host and the storage
controller.

 The current method is to run a stress test, usually a
benchmark utility and output the result we get
This is not sufficient because in this method you do not know
the cause of the good/bad performance you receive, it can be
any number of issues, the host, the fabric, the storage
controller which can be dealing better with some IO patterns
and less with others
The Idea is to perform a host generated IO that will be
acknowledged immediately by the storage controller, no
processing on the storage controller is performed other than
the acknowledge which return as soon as the IO is received .
This will enables us to report to the host the maximal IOPS
that it can expect to get from the machine, given the current
host-fabric-storage configuration
In this case whatever results we get from a benchmark can be
then compared to this number and we can see if the host
reaches it maximal bandwidth or we still need some tweaking in
the host-fabric-storage configuration.

Also this will eliminate the scenario where the customer
claims storage to be slow where in fact he has fabric issues
The first requirement is for the storage controller to be able
to respond immediately for a specific type of IOs without
passing them to any other "head" or "disk module" which will
cost us in latency for the IOs, In the case of the IBM-XIV
storage controller the storage controller will return zeroes
when a host tries to read an area...