rfc3347.txt

一个学习iSCSI协议的文档

      be addressed.

   From section 9.2 Firewalls and Proxy Servers

      SHOULD allow deployment where functional and optimizing middle-
      boxes such as firewalls, proxy servers and NATs are present.

      use of IP addresses and TCP ports SHOULD be firewall friendly.

   From section 9.3 Congestion Control and Transport Selection

      MUST be a good network citizen with TCP-compatible congestion
      control (as defined in [RFC2914]).

      iSCSI implementations MUST NOT use multiple connections as a means
      to avoid transport-layer congestion control.

3. iSCSI Design Considerations

3.1. General Discussion

   Traditionally, storage controllers (e.g., disk array controllers,
   tape library controllers) have supported the SCSI-3 protocol and have
   been attached to computers by SCSI parallel bus or Fibre Channel.



Krueger, et al.              Informational                      [Page 7]

RFC 3347      iSCSI Requirements and Design Considerations     July 2002


   The IP infrastructure offers compelling advantages for volume/
   block-oriented storage attachment.  It offers the opportunity to take
   advantage of the performance/cost benefits provided by competition in
   the Internet marketplace.  This could reduce the cost of storage
   network infrastructure by providing economies arising from the need
   to install and operate only a single type of network.

   In addition, the IP protocol suite offers the opportunity for a rich
   array of management, security and QoS solutions.  Organizations may
   initially choose to operate storage networks based on iSCSI that are
   independent of (isolated from) their current data networks except for
   secure routing of storage management traffic.  These organizations
   anticipated benefits from the high performance/cost of IP equipment
   and the opportunity for a unified management architecture.  As
   security and QoS evolve, it becomes reasonable to build combined
   networks with shared infrastructure; nevertheless, it is likely that
   sophisticated users will choose to keep their storage sub-networks
   isolated to afford the best control of security and QoS to ensure a
   high-performance environment tuned to storage traffic.

   Mapping SCSI over IP also provides:

      -- Extended distance ranges
      -- Connectivity to "carrier class" services that support IP

   The following applications for iSCSI are contemplated:

      -- Local storage access, consolidation, clustering and pooling (as
         in the data center)
      -- Network client access to remote storage (eg. a "storage service
         provider")
      -- Local and remote synchronous and asynchronous mirroring between
         storage controllers
      -- Local and remote backup and recovery

   iSCSI will support the following topologies:

      -- Point-to-point direct connections
      -- Dedicated storage LAN, consisting of one or more LAN segments
      -- Shared LAN, carrying a mix of traditional LAN traffic plus
         storage traffic
      -- LAN-to-WAN extension using IP routers or carrier-provided "IP
         Datatone"
      -- Private networks and the public Internet

   IP LAN-WAN routers may be used to extend the IP storage network to
   the wide area, permitting remote disk access (as for a storage
   utility), synchronous and asynchronous remote mirroring, and remote



Krueger, et al.              Informational                      [Page 8]

RFC 3347      iSCSI Requirements and Design Considerations     July 2002


   backup and restore (as for tape vaulting).  In the WAN,  using TCP
   end-to-end avoids the need for specialized equipment for protocol
   conversion, ensures data reliability, copes with network congestion,
   and provides retransmission strategies adapted to WAN delays.

   The iSCSI technology deployment will involve the following elements:

   (1)  Conclusion of a complete protocol standard and supporting
        implementations;
   (2)  Development of Ethernet storage NICs and related driver and
        protocol software; [NOTE: high-speed applications of iSCSI are
        expected to require significant portions of the iSCSI/TCP/IP
        implementation in hardware to achieve the necessary throughput.]
   (3)  Development of compatible storage controllers; and
   (4)  The likely development of translating gateways to provide
        connectivity between the Ethernet storage network and the Fibre
        Channel and/or parallel-bus SCSI domains.
   (5)  Development of specifications for iSCSI device management such
        as MIBs, LDAP or XML schemas, etc.
   (6)  Development of management and directory service applications to
        support a robust SAN infrastructure.

   Products could initially be offered for Gigabit Ethernet attachment,
   with rapid migration to 10 GbE.  For performance competitive with
   alternative SCSI transports, it will be necessary to implement the
   performance path of the full protocol stack in hardware.  These new
   storage NICs might perform full-stack processing of a complete SCSI
   task, analogous to today's SCSI and Fibre Channel HBAs, and might
   also support all host protocols that use TCP (NFS, CIFS, HTTP, etc).

   The charter of the IETF IP Storage Working Group (IPSWG) describes
   the broad goal of mapping SCSI to IP using a transport that has
   proven congestion avoidance behavior and broad implementation on a
   variety of platforms.  Within that broad charter, several transport
   alternatives may be considered.  Initial IPS work focuses on TCP, and
   this requirements document is restricted to that domain of interest.

3.2. Performance/Cost

   In general, iSCSI MUST allow implementations to equal or improve on
   the current state of the art for SCSI interconnects.  This goal
   breaks down into several types of requirement:

   Cost competitive with alternative storage network technologies:

   In order to be adopted by vendors and the user community, the iSCSI
   protocol MUST enable cost competitive implementations when compared
   to other SCSI transports (Fibre Channel).



Krueger, et al.              Informational                      [Page 9]

RFC 3347      iSCSI Requirements and Design Considerations     July 2002


   Low delay communication:

   Conventional storage access is of a stop-and-wait remote procedure
   call type.  Applications typically employ very little pipelining of
   their storage accesses, and so storage access delay directly impacts
   performance.  The delay imposed by current storage interconnects,
   including protocol processing, is generally in the range of 100
   microseconds.  The use of caching in storage controllers means that
   many storage accesses complete almost instantly, and so the delay of
   the interconnect can have a high relative impact on overall
   performance.  When stop-and-wait IO is used, the delay of the
   interconnect will affect performance.  The iSCSI protocol SHOULD
   minimize control overhead, which adds to delay.

   Low host CPU utilization, equal to or better than current technology:

   For competitive performance, the iSCSI protocol MUST allow three key
   implementation goals to be realized:

   (1)  iSCSI MUST make it possible to build I/O adapters that handle an
        entire SCSI task, as alternative SCSI transport implementations
        do.
   (2)  The protocol SHOULD permit direct data placement ("zero-copy"
        memory architectures, where the I/O adapter reads or writes host
        memory exactly once per disk transaction.
   (3)  The protocol SHOULD NOT impose complex operations on the host
        software, which would increase host instruction path length
        relative to alternatives.

   Direct data placement (zero-copy iSCSI):

   Direct data placement refers to iSCSI data being placed directly "off
   the wire" into the allocated location in memory with no intermediate
   copies.  Direct data placement significantly reduces the memory bus
   and I/O bus loading in the endpoint systems, allowing improved
   performance.  It reduces the memory required for NICs, possibly
   reducing the cost of these solutions.

   This is an important implementation goal.  In an iSCSI system, each
   of the end nodes (for example host computer and storage controller)
   should have ample memory, but the intervening nodes (NIC, switches)
   typically will not.









Krueger, et al.              Informational                     [Page 10]

RFC 3347      iSCSI Requirements and Design Considerations     July 2002


   High bandwidth, bandwidth aggregation:

   The bandwidth (transfer rate, MB/sec) supported by storage
   controllers is rapidly increasing, due to several factors:

      1. Increase in disk spindle and controller performance;
      2. Use of ever-larger caches, and improved caching algorithms;
      3. Increased scale of storage controllers (number of supported
         spindles, speed of interconnects).

   The iSCSI protocol MUST provide for full utilization of available
   link bandwidth.  The protocol MUST also allow an implementation to
   exploit parallelism (multiple connections) at the device interfaces
   and within the interconnect fabric.

   The next two sections further discuss the need for direct data
   placement and high bandwidth.

3.3. Framing

   Framing refers to the addition of information in a header, or the
   data stream to allow implementations to locate the boundaries of an
   iSCSI protocol data unit (PDU) within the TCP byte stream.  There are
   two technical requirements driving framing: interfacing needs, and
   accelerated processing needs.

   A framing solution that addresses the "interfacing needs" of the
   iSCSI protocol will facilitate the implementation of a message-based
   upper layer protocol (iSCSI) on top of an underlying byte streaming
   protocol (TCP).  Since TCP is a reliable transport, this can be
   accomplished by including a length field in the iSCSI header. Finding
   the protocol frame assumes that the receiver will parse from the
   beginning of the TCP data stream, and never make a mistake (lose
   alignment on packet headers).

   The other technical requirement for framing, "accelerated
   processing", stems from the need to handle increasingly higher data
   rates in the physical media interface.  Two needs arise from higher
   data rates:

   (1)  LAN environment - NIC vendors seek ways to provide "zero-copy"
        methods of moving data directly from the wire into application
        buffers.

   (2)  WAN environment- the emergence of high bandwidth, high latency,
        low bit error rate physical media places huge buffer
        requirements on the physical interface solutions.




Krueger, et al.              Informational                     [Page 11]

RFC 3347      iSCSI Requirements and Design Considerations     July 2002


   First, vendors are producing network processing hardware that
   offloads network protocols to hardware solutions to achieve higher
   data rates.  The concept of "zero-copy" seeks to store blocks of data
   in appropriate memory locations (aligned) directly off the wire, even
   when data is reordered due to packet loss.  This is necessary to
   drive actual data rates of 10 Gigabit/sec and beyond.

   Secondly, in order for iSCSI to be successful in the WAN arena it
   must be possible to operate efficiently in high bandwidth, high delay
   networks.  The emergence of multi-gigabit IP networks with latencies
   in the tens to hundreds of milliseconds presents a challenge.  To
   fill such large pipes, it is necessary to have tens of megabytes of
   outstanding requests from the application.  In addition, some
   protocols potentially require tens of megabytes at the transport
   layer to deal with buffering for reassembly of data when packets are
   received out-of-order.

   In both cases, the issue is the desire to minimize the amount of
   memory and memory bandwidth required for iSCSI hardware solutions.

   Consider that a network pipe at 10 Gbps x 200 msec holds 250 MB.
   [Assume land-based communication with a spot half way around the
   world at the equator.  Ignore additional distance due to cable
   routing.  Ignore repeater and switching delays; consider only a
   speed-of-light delay of 5 microsec/km.  The circumference of the
   globe at the equator is approx. 40000 km (round-trip delay must be
   considered to keep the pipe full).  10 Gb/sec x 40000 km x 5
   microsec/km x B / 8b = 250 MB].  In a conventional TCP
   implementation, loss of a TCP segment means that stream processing
   MUST stop until that segment is recovered, which takes at least a
   time of <network round trip> to accomplish.  Following the example
   above, an implementation would be obliged to catch 250 MB of data
   into an anonymous buffer before resuming stream processing; later,
   this data would need to be moved to its proper location.  Some
   proponents of iSCSI seek some means of putting data directly where it
   belongs, and avoiding extra data movement in the case of segment
   drop.  This is a key concept in understanding the debate behind
   framing methodologies.

   The framing of the iSCSI protocol impacts both the "interfacing
   needs" and the "accelerated processing needs", however, while
   including a length in a header may suffice for the "interfacing
   needs", it will not serve the direct data placement needs.  The
   framing mechanism developed should allow resynchronization of packet
   boundaries even in the case where a packet is temporarily missing in
   the incoming data stream.





Krueger, et al.              Informational                     [Page 12]

RFC 3347      iSCSI Requirements and Design Considerations     July 2002


3.4. High bandwidth, bandwidth aggregation

   At today's block storage transport throughput, any single link can be
   saturated by the volume of storage traffic.  Scientific data
   applications and data replication are examples of storage
   applications that push the limits of throughput.

   Some applications, such as log updates, streaming tape, and
   replication, require ordering of updates and thus ordering of SCSI
   commands.  An initiator may maintain ordering by waiting for each
   update to complete before issuing the next (a.k.a. synchronous
   updates).  However, the throughput of synchronous updates decreases
   inversely with increases in network distances.

   For greater throughput, the SCSI task queuing mechanism allows an
   initiator to have multiple commands outstanding at the target
   simultaneously and to express ordering constraints on the execution
   of those commands.  The task queuing mechanism is only effective if
   the commands arrive at the target in the order they were presented to
   the initiator (FIFO order).  The iSCSI standard must provide an
   ordered transport of SCSI commands, even when commands are sent along
   different network paths (see Section 5.2 SCSI).  This is referred to
   as "command ordering".

   The iSCSI protocol MUST operate over a single TCP connection to
   accommodate lower cost implementations.  To enable higher performance
   storage devices, the protocol should specify a means to allow
   operation over multiple connections while maintaining the behavior of
   a single SCSI port.  This would allow the initiator and target to use
   multiple network interfaces and multiple paths through the network
   for increased throughput.  There are a few potential ways to satisfy
   the multiple path and ordering requirements.

   A popular way to satisfy the multiple-path requirement is to have a
   driver above the SCSI layer instantiate multiple copies of the SCSI
   transport, each communicating to the target along a different path.
   "Wedge" drivers use this technique today to attain high performance.
   Unfortunately, wedge drivers must wait for acknowledgement of
   completion of each request (stop-and-wait) to ensure ordered updates.

   Another approach might be for iSCSI protocol to use multiple
   instances of its underlying transport (e.g. TCP).  The iSCSI layer
   would make these independent transport instances appear as one SCSI
   transport instance and maintain the ability to do ordered SCSI
   command queuing.  The document will refer to this technique as
   "connection binding" for convenience.





Krueger, et al.              Informational                     [Page 13]