Cisco Cisco MDS 9500 Series Supervisor-2 Module Libro blanco

Important notices, privacy statements, and trademarks of Cisco Systems, Inc. can be found on cisco.com.

Page 3 of 16

Device Performance and Oversubscription Ratios

Oversubscription is a necessity of any networked infrastructure and directly relates to the major benefit of a network—to share common resources

among numerous clients. The higher the rate of oversubscription, the lower the cost of the underlying network infrastructure and shared resources.

Because storage subsystem I/O resources are not commonly consumed at 100 percent all the time by a single client, a fan-out ratio of storage

subsystem ports can be achieved based on the I/O demands of various applications and server platforms. Most major disk subsystem vendors provide

guidelines as to the recommended fan-out ratio of subsystem client-side ports to server connections. These recommendations are often in the range

of 7:1 to 15:1.

When considering all the performance characteristics of the SAN infrastructure and the servers and storage devices, two oversubscription metrics

must be managed: IOPS and network bandwidth capacity of the SAN. The two metrics are closely related, although they pertain to different

elements of the SAN. IOPS performance relates only to the servers and storage devices and their ability to handle high numbers of I/O operations,

whereas bandwidth capacity relates to all devices in the SAN, including the SAN infrastructure itself. On the server side, the required bandwidth is

strictly derived from the I/O load, which is derived from factors including I/O size, percentage of reads versus writes, CPU capacity, application I/O

requests, and I/O service time from the target device. On the storage side, the supported bandwidth is again strictly derived from the IOPS capacity

of the disk subsystem itself, including the system architecture, cache, disk controllers, and actual disks.

In most cases, neither application server host bus adapters (HBAs) nor disk subsystem client-side controllers are able to handle full wire-rate

sustained bandwidth. Although ideal scenario tests can be contrived using larger I/Os, large CPUs, and sequential I/O operations to show wire-rate

performance, this is far from a practical real-world implementation. In more common scenarios, I/O composition, server-side resources, and

application I/O patterns do not result in sustained full-bandwidth utilization. Because of this fact, oversubscription can be safely factored into

SAN design. However, you must account for burst I/O traffic, which might temporarily require high-rate I/O service. The general principle in

optimizing design oversubscription is to group applications or servers that burst high I/O rates at different time slots within the daily production

cycle. This grouping can examine either complementary application I/O profiles or careful scheduling of I/O-intensive activities such as backups

and batch jobs. In this case, peak time I/O traffic contention is minimized, and the SAN design oversubscription has little effect on I/O contention.

Best-practice would be to build a SAN design using a topology that derives a relatively conservative oversubscription ratio (for example, 8:1)

coupled with monitoring of the traffic on the switch ports connected to storage arrays and Inter-Switch Links (ISLs) to see if bandwidth is a limiting

factor. If bandwidth is not the limited factor, application server performance is acceptable, and application performance can be monitored closely,

the oversubscription ratio can be increased gradually to a level that is both maximizing performance while minimizing cost.

Traffic Management

Are there any differing performance requirements for different application servers? Should bandwidth be reserved or preference be given to traffic

in the case of congestion? Given two alternate traffic paths between data centers with differing distances, should traffic use one path in preference

to the other?

For some SAN designs it makes sense to implement traffic management policies that influence traffic flow and relative traffic priorities.

Fault Isolation

Consolidating multiple areas of storage into a single physical fabric both increases storage utilization and reduces the administrative overhead

associated with centralized storage management. The major drawback is that faults are no longer isolated within individual storage areas.

Many organizations would like to consolidate their storage infrastructure into a single physical fabric, but both technical and business

challenges make this difficult.