Brocade's switch earns big numbers
 
|
|||
 | |||
|
|
|
|||
| |||
|
If you build a bigger switch, they will come.
In a contest between two of only three shipping high-port-density, director-class storage-area network switches today, our Blue Ribbon Award goes to Brocade Communications Systems' SilkWorm 12000, primarily on the strength of its showing in our configuration and performance test categories.
McData's Intrepid 6000 Series 64-Port Director is a solid product with some impressive high availability and management tools. Inrange's FC/9000 is the third director-class switch, but the company declined to participate.
Director switches are high-port-count (64 ports and higher), high-availability SAN switches designed with one principal goal - to minimize hops in a SAN. While minimizing hops in any network is a good idea, it carries a premium in storage networks, which transport delay-sensitive disk I/Os that must arrive in order.
McData fielded the original director switch to interconnect IBM S/390 mainframes with target storage devices over Enterprise Systems Connection (ESCON) networks. When Fibre Channel SANs presented opportunities in the Unix/LAN world, McData ported its director switch know-how into that market.
Brocade, a dominant presence in the Unix/LAN SAN switch world, has been criticized for being considerably late to market with its director-class SilkWorm 12000, which we tested just two weeks before its March release date.
Brocade says it made a strategic decision to wait to ship its product until it included full 2G bit/sec throughput support. We verified the SilkWorm's 2G bit/sec transport level when it posted an impressive full-duplex maximum throughput rate of almost 210M byte/sec. But since it is still a 1G bit/sec world, and in the interest of effective comparison, we also tested the Brocade box at 1G bit/sec
Using SANs for business continuity
How we did it
Bandwidth distribution charts
The feature presentation
Interactive Scorecard and NetResults
Archive of Network World reviews
Subscribe to the Product Review newsletter
The Brocade product was also a winner in our configuration category. Brocade offered legacy support for Arbitrated Loop devices and hardware zoning. With 128 maximum ports, SilkWorm offered the ability to deploy two discrete 64-port director switches within the same 14-unit chassis. Both products supported redundant CPUs.
However, McData had two important redundancy advantages. The first was a hot-standby backplane. The second was the fact that Intrepid's 64 maximum ports were interspersed throughout 16 chassis slots, while the same number of ports fit in only four slots on SilkWorm. Greater per-module port density is beneficial from a real estate standpoint, but failure of a single module on SilkWorm will ruin your day faster than the same occurrence on Intrepid.
Directors as performers
We tested each box in five performance areas.
For our partial-mesh throughput testing, we used Spirent Communications' SmartBits to configure 32 initiator ports to send the largest (2,148 bytes) and smallest (64 bytes) Fibre Channel frames in separate tests to 32 target ports in a one-port-to-one-port configuration (see How we did it).
Brocade performed at line rate (105.03M and 77.27M byte/sec, respectively) with large and small frames through one director, and very nearly the same in a two-hop environment (see graphic, page 74). McData's Intrepid 6000 performed well enough with large frames, but was clearly less efficient at handling smaller frames, delivering 37.72M and 34.99M byte/sec in a one- and two-hop environment, respectively. While McData engineers noted that the overwhelming majority of Fibre Channel frames are more than 400 bytes, the ability to process large and small frames should not be mutually exclusive propositions.
With our high-stress, full-mesh throughput test, we configured each of 64 SmartBits ports to send large frames to, and receive from, each of the other 63 ports. Again, SilkWorm demonstrated impressive efficiency by sustaining an average per-port throughput rate of 104.37M byte/sec, compared with Intrepid's 84.51M byte/sec.
We also tested for head-of-line blocking, which analyzes a switch's ability to sustain performance despite congestion on any one port. We split full line-rate traffic from one SmartBits initiator port to each of two ports on the director switches. Half of the traffic, or approximately 50M byte/sec, targeted one port (Port A), and the rest was bound for another port (Port B). With a second initiator SmartBits port, we sent 100% line rate to Port A. The idea was to observe how the congestion on Port A affected the uncongested Port B.
All traffic bound for Port B on the SilkWorm 12000 arrived at precisely the offered rate with no appreciable delay. SilkWorm's flow control logic throttled the link offering 100M byte/sec down to 50M byte/sec, resulting in an even allocation of bandwidth across the three data streams with no impedance on the uncongested port.
Intrepid handled things differently. The uncongested port on Intrepid realized only 20M byte/sec out of the 50M byte/
sec offered. McData's device gave priority to the 100M byte/sec source port. To make that possible, it also had to push back on the other source port, resulting in an 80-20-20M byte/sec bandwidth distribution across the three streams. This result impeded throughput to the uncongested port and on the switch overall.
McData explained that Intrepid was engineered to give priority to "hot" ports, reasoning that without a de facto Fibre Channel quality-of-service standard in place, this is as valid a method of prioritizing traffic through a SAN switch as any. Given the confines of Fibre Channel's "stop-or-go" method of flow control, this argument is compelling and likely to play well in the Fibre Channel community. However, we find no precedent in any Layer 2 switching technology for taxing uncongested ports to benefit congested ones.
With the 15-to-1 throughput test, we sought to observe how each device managed an extreme case of many servers requiring access to the same target storage port. The results show that Intrepid 6000 distributed its bandwidth more evenly among the servers. It scored a throughput range of 4.72M to 9.85M byte/sec across all 15 servers, while Brocade's distribution range was broader, at 3.72M to 14.86M byte/sec across the servers.
But bandwidth distribution testing across interswitch links decidedly favored SilkWorm. The object of this test was to see how the switches distributed three ports worth of traffic through only two trunk links. The data points for Brocade are virtually flat (51.81M to 52.06M byte/sec range), while McData's distribution across the links is more scattered, ranging from 27.24M to 40.5M byte/sec.
While both devices were strong in terms of our availability testing, the Intrepid showing was a little better. McData's CPU failed over in 500 msec, while Brocade's CPU took a full 18 seconds. The McData product also offers a nondisruptive code load and activation feature. Using this feature, we upgraded the firmware on an active switch without interruption of service.
McData's power supplies and redundant crossbar backplane failover worked as advertised. Brocade demonstrated its 2N power supply redundancy without a hitch.
In our reboot test, we simply cut off all power to the devices to see how quickly the devices could resume normal operation. Intrepid resumed service in 1 minute, 58 seconds; while Silkworm was back online in 2 minutes, 36 seconds.
To test component failure resistance, we pulled active modules out of the switch randomly. With only one port module and one CPU module left, both devices functioned normally. Redundant interswitch links on both devices failed over seamlessly.
Managing the director
In our experience, SAN switches as a species have supported well-conceived management interfaces effectively, and the two director switches we tested were no exception. One recurrent complaint about Fibre Channel SANs is the required parallel management of a separate network that is replete with foreign terms and concepts.
Both products addressed our management and administration criteria by straddling two software products - one for high-level, multidevice management; the other for local drill-down into the switches themselves.
McData led with SANavigator, Version 2.7, which it acquired from Western Digital in 2001 to compete with multivendor SAN management software packages such as Veritas Software's SANPoint and Tivoli's Storage Manager. SANavigator is a client/
server-based package that features a slick, autodiscovered topology map of the existing SAN, complete with color-coded traffic pattern and alarm animations. Software links let you double-click on any icon to launch a Telnet or a browser session with the device it represents.
Double-clicking on the Intrepid icon launched its local, browser-based management tool called Enterprise Fabric Connectivity (EFC) Manager that offers a detailed image of the device. While SANavigator is well-organized, we found a dead end that required closing out the program and relaunching to get back to the main interface. Real-time monitoring of Intrepid meant drilling down with EFC Manager. It offered lots of real-time counters, but the need to manually refresh the counters was annoying, as was the lack of environmental statistics.
While both applications supported SNMP with all the right Management Information Bases, neither supported in-band notification of any kind, requiring an IP network for all visibility into the SAN. But the two packages do have some handy notification mechanisms. SANavigator has a software add-on that will send notifications to your PDA; EFC Manager's "call home" function will send text messages to your favorite network operations center over a modem connection. Both can send Simple Mail Transfer Protocol-based notifications.
Both tools had audit logs, giving IT administrators in large shops the knowledge of who performed what configuration changes, when, and from where. SANavigator also included a hardware log that was essentially a learned inventory of the SAN's elements, complete with vendor name and serial number of the device. SANavigator included several useful canned reports that can be generated and viewed via browser or SQL database client, e-mailed or printed.
For proactive troubleshooting, EFC Manager users can perform loop-back tests for card modules and individual Gigabit interface cards. It also allowed "beaconing," or the ability to ping the LED on a given port for physical identification. Also laudable was EFC Manager's granular permission levels, which let large shops manage levels of access on a per-user basis.
Brocade's Fabric Manager provides management access to multiple switches on multiple switch fabrics, although it doesn't offer the same high-level view as SANavigator. Because it could only manage Brocade products, it offered no context of the surrounding SAN that the switches were powering.
One effective tool in the suite is Performance Monitor, a software module that features chart-based, real-time, granular statistics of any managed switch. These charts could be printed for reporting purposes, but they could not be handled electronically in any form. It offers a variety of real-time, per-port counters. Like EFC Manager, its base for navigating the program's screens was a graphical representation of the device, complete with LEDs blinking in real time.
Easily the strongest aspect of Brocade's management is its variety of powerful in-band and out-of-band notification mechanisms. Servers outfitted with host bus adapters, or Fibre Channel network interface cards, could enjoy SNMP, Syslog, Telnet or HTTP-based information about any switch either in-band via Fibre Channel over IP; or out-of-band via an IP network. What's more, in a multiple-switch SAN, Brocade switches can funnel messages back through the Fibre Channel network into one switch defined as the "management gateway," which will forward the information through its LAN connection.
WebTools offers adequate logging. The entries ranged from reasonably descriptive to cryptic, and the Help function was not helpful in clarifying these. Outside of printing reports from Performance Monitor, reporting on the Brocade management tools was anemic, offering a few marginally useful reports. Brocade offers an assortment of useful diagnostic commands from its main command line, but only beaconing was available on either of the graphical user interfaces themselves.
Although the outcome of these tests indicates some clear competitive advantages for Brocade, McData's Intrepid 6000 very likely has the goods to retain much of its entrenched incumbency.
 |
|||||||||||||||||||
|
|||||||||||||||||||
|
|||||||||||||||||||
| |||||||||||||||||||
Back to top
RELATED LINKS
Percy is a technology analyst at Miercom, an independent testing lab located in Princeton Junction, N.J. He can be reached at kpercy@mier.com.
Percy is also a member of the Network World Global Test Alliance, a cooperative of the premier reviewers in the network industry, each bringing to bear years of practical experience on every review. For more Test Alliance information, including what it takes to become a member, go to www.nwfusion.com/alliance.
Review archive
Links to all of our past reviews.
The skinny on iSCSI SANs
IP storage products prove they can talk to each other and to Fibre Channel gear. Network World, 05/06/02.
Panel: Storage options improving
Storage experts say IP storage and performance advances present a variety of opportunities for enterprise customers who have not yet implemented storage-area networks. Network World, 01/21/02.
Audio primer: Storage-area networks
Unlike stand-alone servers, a storage area network distributes resources over a scalable, redundant network of servers and storage devices. Network World Fusion.
