Fibre channel arbitrated loop bufferless switch circuitry to...

Details Fibre channel arbitrated loop bufferless switch circuitry to... Fibre channel arbitrated loop bufferless switch circuitry to...

1998-11-19

2003-09-02

Nguyen, Steven (Department: 2665)

Multiplex communications

Pathfinding or routing

Switching a message which includes an address header

C370S230000, C370S535000, C370S411000, C370S462000

Reexamination Certificate

active

06614796

ABSTRACT:

FIELD OF USE
Fibre Channel networks are known loop configuration networks that have a plurality of known type nodes such as servers, printers, disk arrays etc. all connected together by the loop. Such networks use a unique protocol involving a plurality of 40 bit primitives that are used to arbitrate for loop control, to establish connections and to carry out flow control for data transfers of frames of data. The flow control inherent to the Fibre Channel Arbitrated Loop network (hereafter FCAL nets) protocol has the advantage of eliminating the need for the nodes to have extensive buffering capabilities since the destination node controls the amount of data it receives by transmission of an RRDY primitive to the source node each time the destination node is ready to: receive another frame.
Fibre Channel networks emerged as a family of interconnection topologies to increase bandwidth over fast-wide SCSI networks and to increase the number of server and storage elements that can be connected to 126 over the 16 device limit of SCSI. Advantages of FCAL include that devices may be farther apart (up to 10 km) and more numerous and that the size of data transfers is very large compared to the overhead that is required to set up every transfer. This makes FCAL very efficient and more attractive than less efficient protocols such as TCP/IP over Ethernet and SCSI over a bus connection.
Hub based network topologies are generally desirable because they overcome certain limitations on the number of nodes that can be coupled to a network by breaking it up into segments coupled by the hub. Many Ethernet networks use hubs as do token ring networks. Hubs in FCAL networks receive packets from a source node on an input line coupled to the source node and rebroadcast the packet on an output line coupled to the next node which rebroadcasts the packet to the next node and so on. The rebroadcast by subsequent nodes in the chain wastes computing resources. Switched topologies work differently in that packets are not rebroadcast, but instead are connected directly to the line coupled to the destination node thereby eliminating processing by other nodes which are not the destination to receive and rebroadcast messages not destined for that node.
Despite their advantages, a significant problem in FCAL networks is delay and this delay increases as the network scales up in size. Each meter of cable contributes 5 ns of delay. Further, each node contains an elasticity buffer or FIFO to absorb the differences between incoming and outgoing data rates. Data passing through a node enroute to its destination passes through the nodes elasticity buffer and suffers a typical delay of 3 words. Typically, disk clusters are 10 drives to a cluster with each drive being one node and imposing its own delay. If there are 10 clusters coupled to a server, this would represent typically 5.3 microseconds of delay in transition of each primitive and data frame travelling around the loop. In other words, this delay is imposed on each loop tenancy. In an I/O operation, there are typically 4 tenancies for a write to disk, each involving 3 “round trips”: ARB, OPN-RRDY and Data/CLS (see ANSI standard X3T10FCP which is hereby incorporated by reference). Thus, 12 delays would be suffered by each command transaction. On a 100 node loop, this translates to approximately 64 microseconds of delay per command.
“The command overhead of modem disk drives is around 200 microseconds and falling. The delay per command coupled to the command overhead of the drive imposes a significant penalty on performance of approximately 32%. For random access benchmarks with small I/O payloads typical of database queries, the performance penalty becomes more pronounced. The problem manifests itself as the inability of the server to achieve more I/O operations per second, regardless of how many more disk drives are added to the system.”
Spatial reuse provided by switches or hubs which allow concurrent loop tenancies is one way of reducing the delay problem. The IBM serial storage architecture in the prior art is one method of providing spatial reuse.
Connection oriented switched topologies were tried in early Fibre Channel Fabric networks to attempt to overcome the delay problems of loops by cutting down the number of nodes each primitive and data frame passes through in getting from source to destination and providing spatial reuse. These early fabric switches were complicated, expensive and slow, all of these characteristics being found quite undesirable by artisans of FCAL networks. In the early FC Fabric switches, an entire frame of data with a header that indicated the destination node to which the frame was directed was sent to the switch for purposes of requesting a connection. These early switch designs had microprocessors which were used to implement several layers of software architecture to receive the frame, pass it up through various layers of processing to find the frame boundaries, crack the frame open, determine its destination address and then attempt to find the destination node and make the switching connection. The entire frame of data had to be buffered during this process of attempting to find the destination and make the proper connection. It was possible in this early design that the connection was never made, because, for example, the destination node was busy with another conversation. The switch would then have to send a message back to the source that no connection was made and to try again later. Because of limited buffer space in the switch, the data in the original frame might need to be overwritten by other data from a frame of data embodying another request. In such a case, the switch would have to send another message to the source saying, “Sorry, I lost your data. Execute error recovery protocol.” Error recovery protocols further complicated the operation and construction of such systems. If a connection is made, the switch receives another frame of data back from the destination. This frame also must be received, have its boundaries detected and must be cracked open to examine its contents to see if the destination is saying, “Yes, I am available for a connection.” This type of switch proved to be unworkable and FCAL loops became the standard interconnect for disks and servers.
Prior art Fibre Channel switches are commercially available from Ancor and Brocade Communications which provide spatial reuse and efficient link utilization. The FL_ports connected to these switches also address physical delays as they pertain to FCAL. However, these switches require link rate frame buffering to accomplish their performance levels, and also operate on the entire 24-bit address contained in the FC frame. In contrast, the invention described herein uses zero buffering and an 8-bit address decode for a much more efficient and inexpensive design.
Many network switched topologies that use entire frames of data to request a connection through the switch suffer these same drawbacks. The need for errory recovery protocols arise because of the potential for lost data arising from the fact that only limited amount of memory can be put in the switch at realistic costs, and in heavy traffic situations, the memory may be exhausted and some portion thereof may have to be rewritten with new data before the original data is delivered. Memory is expensive, takes up space and complicates the design.
Examples of other network topologies other than Fibre Channel Fabric that suffer these same drawbacks are the 1 Gigabit Ethernet® and ATM protocol networks now in public use.
The Fibre Channel Arbitrated Loop (FCAL) topology emerged as a way of providing simple, low-cost connectivity to more nodes over a shared media than could be provided in point-to-point topologies without the requirement for an expensive fabric switch. FCAL networks allow up to 126 node ports to be coupled by a shared media using a simple protocol without the need for a separate fabric switch. Unlike the switched fabric topology which has a centralized approach to routing, FCAL netw

Affiliated with

Also associated with

Rating

Fibre channel arbitrated loop bufferless switch circuitry to... does not yet have a rating. At this time, there are no reviews or comments for this patent.

If you have personal experience with Fibre channel arbitrated loop bufferless switch circuitry to..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Fibre channel arbitrated loop bufferless switch circuitry to... will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFUS-PAI-O-3058272