In a May 2000 paper issued soon after Nishan promised to revolutionize (yawn) the world with Storage over IP, the Evaluator Group, whose principals collectively have over 200 man years in the storage business, basically dismantled the arguments of the Storage over IP crowd. To this date, there has been no response except vague mumblings of stealth crash programs and contradictory rhetoric that I swear, come this close to implying that business applications will be dumbed down to make the procrustean fit.
Here are some excerpts:
SCSI over IP
The Problem Attempting to be Solved
The question that immediately is asked is "Why do SCSI over IP?" This question is asked because Storage Area Networks are being implemented with fibre channel infrastructures today with multitudes of companies offering solutions of some type. As usual, there’s no one answer but there are a few that are often quoted.
The first is that some people have a vision of having a single network for everything. Everything in this case includes storage, as well as the typical Ethernet functions as email, FTP of data, web access, device management over SNMP (even management of traditional non-data processing appliances), voice and video transmission, and all the other uses. These visionaries cite the usage of a single
infrastructure, massive network bandwidth to be achieved in the future, and the ability to have storage anywhere in the world with the user (and even the administrator) not needing to know or worry about where........
[If information is power and the cost of one hour of downtime for an online brokerage can cost up to $6.5M, what is wrong with this one-size-fits-all picture?]
Some of the other issues that confront SCSI over IP are:
* The implementation of IP has a significant amount of software that must be executed by the initiating server. This software has many different elements including the requirement for checking for lost packets, interrupt handling on each packet transmitted, integrity check generation and verification, buffer movement of data, etc. In a fibre channel implementation, this has all been
implemented in hardware. There is some improvement that can be made for IP with implementing some of the protocol stack in hardware but that will not eliminate the software overhead content. This software overhead is one of the contributors in the disparity in performance between accesses that are done at a block level over fibre channel and the file level access over Ethernet. As an example, measurements have been performed that show the effective data rate of Gigabit Ethernet when data is being moved to be
27MB/s maximum whereas fibre channel has been demonstrated at greater than 90MB/s with the 1GHz implementation.
* Congestion on IP networks leads to some very impacting consequences. Since the solution to congestion on Ethernet is to drop packets, the overhead will actually increase because it is up to the receiver to detect that a packet
has not arrived and then to request it (over the same Ethernet) to be sent again. This creates significant network performance problems that are typically called meltdown. Storage devices have a significant problem if
data does not arrive – especially in the case of a missing packet. In fibre channel, storage devices generally implement class 2 service, which means guaranteed delivery. The sequence checking and acknowledgement is handled in hardware so that data is guaranteed to be at the storage device or from the storage device without incurring additional network overhead.
* Most corporate environments require a level of performance from their storage systems that is a major element in their business. This requires a consistent, guaranteed level of performance from the storage subsystem. IP is unable to provide a guaranteed performance with the varying overhead and congestion problems noted earlier.
* One criticism that must be made is that most of the people making the case for SCSI over IP do not have storage backgrounds – backgrounds in the development or detailed understanding of storage systems as utilized in block I/O for performance critical environments. A network person without this background has less credibility.
[Well, somebody's got to sell the stock]
Related Companion Solutions
There are several additional solutions that may be utilized with SCSI over IP to improve on the performance, which is one of the weaknesses identified.
* As described earlier, implementing part of the protocol stack in hardware can greatly improve the performance by reducing the amount of software overhead required.
· Increasing the speed of the Ethernet to 10Gigabits per second will provide more raw bandwidth. The efficiency of this increase will not change, just the raw bandwidth. It should be noted that corresponding changes in the
competing technology, fibre channel, should be expected as well. With the usage of 10Gb/s, the connections will probably be required to be fiber for the complete connection and may require physical infrastructure changes.
[Earlier this year, the cost disparity between Ethernet (25+ years old tech) and Fibre Channel (10+ years old tech) was roughly 2:1 in favor of Ethernet. That gap is narrowing more quickly than, I think, the everything-on- Ethernet crowd is prepared for. Also, Gigabit Ethernet and Fibre Channel share the same physical layer. The physical layer for 10 Gb/s Gigabit Ethernet DOES NOT currently exist and you know how petty this my-vaporware-is-better-than-your-vaporware discussion can get.]
* Another performance improvement item is to increase the packet size for IP. Currently, the packet size is limited to 1,518 bytes (in reality, most transfers are much smaller – on the order of 128 bytes). Fibre channel has a base packet size of 2,112 bytes but can define a maximum transfer unit sequence of up to 64 frames to allow transfers of up to 128 Megabytes without interrupting the processor and incurring additional overhead.
Other Approaches to Attack the Problem
There have been some other approaches to attack the same problems that SCSI over IP is trying to solve. Obviously, SANs implemented over fibre channel are solving the problem as defined today except for the single network issue (which is not an acceptable proposition for storage professionals). Other solutions include:
* Use the Virtual Interface Architecture instead of IP over Ethernet for storage access. This is targeted at greatly reducing the software overhead compared to IP but requires that the applications that exploit VI must change.
Additionally, the storage system must implement VI. The change to applications may be a significant impediment for existing applications/customers but may be an opportunity for a new implementation.
* InfiniBand is being developed as a replacement for the PCI bus utilized internally in many servers. InfiniBand may grow outside of the server to allow switched, point-to-point access to data direct to the processor. This will take time to evolve and would represent an infrastructure change. This will be several years before the viability as a storage system connection in existing environments is reached.
* IP over fibre channel is one solution that has been developed by host bus adapter vendors. Several adapters are now available that will route IP over fibre channel. The initial target for this capability was the clustering heartbeat function and out-of-band management functions. This may be of limited but valuable functionality however some vendors may try to make it out to be significantly more.
* Use the Virtual Interface Architecture over fibre channel for storage access and clustering controls. Again, this will greatly reduce overhead and the DMA transfers utilized in VI can be mapped directly onto fibre channel. Existing usages will work without modification but new applications that exploit VI must change.
Conclusion
Storage can be done over IP and is being done today with NAS. There are improvements in performance being made with IP but improvements are also being made in fibre channel. In fibre channel, 2Gb/s is being deployed today and 10Gb/s
is in development. The bottom line is that storage over IP is not for everyone and because of the security, integrity, and guaranteed performance concerns. Demanding environments will need direct or SAN attached storage.
There will be coexistence of NAS devices and SANs and even linkage between the two for backup and storage backend for the NAS device in some cases. SCSI over IP is one of those performance improvement areas but does not solve all the
problems and introduces others. This will be developed and made available at some time but will not delay the implementation of fibre channel based SANs.
Companies that are not in the storage business that are trying to make an impact with SCSI over IP are trying to get a share of the market and be successful. Storage professionals will be very discerning when it comes to storage and solutions. The infrastructure implementation for fibre channel based SANs will continue and they will be resistive to make another change in the next 5 to 7 years.
evaluatorgroup.com |
