What I liked most about the presentation is that I finally understand the difference in market size, and the relationship between EMC and NTAP. That is, EMCs market is in 2000 is about $44 billion. The NAS submarket, in which EMC participates but NTAP dominates, is about $2 billion. This gives me a better perspective overall. EMC estimates the Storage Market to be about $77 billion by 2003.
It does show up in the breadth and depth of the solutions that EMC provides which, in turn, shows up in the diversity of EMC's revenue base. Note that EMC's newest products are growing faster -- in absolute terms and in percentage terms -- than NTAP's entire business.
Here are some other key differences that may provide some food for thought to those who persist on the counter-productive endeavor of pitting EMC vs NTAP.
1) RAID level support
EMC pioneered the use of cached RAID in 1990. Here's a history of RAID provided by the Raid Advisory Board that should be an eye-opener to those who believe that both are mutually exclusive:
Two types of RAID provide EDAP (Extended Data Availability and Protection) for disks: Mirroring and Parity RAID. Mirroring predated Parity RAID and was identified in the Berkeley Papers as RAID Level 1. Its disadvantage is that, unlike Parity RAID, Mirroring requires 100% redundancy. Its advantages, unlike Parity RAID, are that read performance is improved, the impact on write performance is generally modest and a higher percentage of disks in a Mirrored redundancy group may fail simultaneously as compared to a Parity RAID redundancy group.
Parity RAID is identified in the Berkeley Papers as RAID Levels 3, 4, 5 and 6. In these cases, overhead (redundant data in the form of Parity) as compared to Mirroring (redundant data in the form of a complete copy) is significantly reduced to a range of 10% to 33%. In all cases, there are varying degrees of performance degradation intrinsically associated with Parity RAID. The differences between RAID Levels 3-6 pertain to the manner in which data and redundant data are mapped to the disks comprising the RAID array. These differences are of concern to RAID designers, but are generally of little interest and importance to end-users.
One exception to this rule is RAID Level 6. Unlike RAID Levels 3-5 which provide EDAP in the event of a failure of one disk, RAID Level 6 provides EDAP when two disks fail at the same time or when a second disk fails during the reconstruction period resulting from the first disk failure.
raid-advisory.com
In 1995, RAB decided that many RAID vendors were monkeying too much with the RAID numerical classification system and decided to implement an overlay classification system on top of the numerical system:
Failure Tolerant Disk System Plus (FTDS+)
Failure Tolerant Disk System Plus Plus (FTDS++)
Disaster Tolerant Disk System Plus (DTDS)
Disaster Tolerant Disk System Plus (DTDS+)
EMC was the first vendor to achieve a DTDS+ rating -- the highest ranking under this more stringent classification scheme in 1997. Even in a clustered failover mode, NTAP's filers currently DO NOT qualify for this ranking. Here are the vendors who have submitted their products for conformance under this new classification:
raid-advisory.com
EMC was also one of the first vendors to support multiple RAID levels in one box in 1995 or 1996 when they entered the open systems market. NTAP is still exclusively a parity RAID vendor (proprietary version of RAID 4). For a vendor hoping to reduce its very high dependence on the revenue-poor/multiple-rich dotcom market (35% to 40%) and penetrate the Global 2000 enterprise market, that is a very serious handicap.
The reason that functionality is important is that the ability to support mirrored and parity RAID directly affects the quality of the software that layers on top of the RAID system, including but not limited to snapshot and mirroring software.
Here's why snapshot and mirroring software are important to Global 2000 company. The typical Global 2000 company that has implemented ERP has a unique problem. It's not unusual for each transaction to generate more than 15 copies of the same data in multiple databases across the enterprise. Obviously, mirroring all those databases is not an economical solution, but storing everything on parity-based RAID arrays does not quite offer the same protection as mirroring either. A Symmetrix or a dedicated network (ESCON or Fibre Channel) of Symmetrix boxes can provide mirroring (with real-time remote mirroring and geographic clustering abilities) and parity RAID in the same box so that allows a company to identify the production databases worth mirroring (with dedicated disk space) and the production databases for which a parity RAID implementation will suffice.
Where else does this handicap show up in NTAP's solutions?
NTAP's snapshot technology can only support asynchronous mirroring (unpredictable latency) while EMC has supported synchronous (predictable latency) mirroring and semi-synchronous (bounded latency) mirroring since 1994 with SRDF and asynchronous (unpredictable latency) mirroring with the introduction of SRDF over IP earlier this year. It is well understood that in asynchronous mirroring, a certain portion of the data is always at risk. That's the structural weakness of NTAP's one-dimensional parity RAID-based snapshot technology. NTAP is working on making its filers support mirroring (RAID 1) next year. Once they are able to do that, their Snapshot technology will be on par with the Unix/NT version of EMC's SRDF and Timefinder technology......on paper. That gives EMC with anywhere from a 3-6 year lead over NTAP in those areas.
2) CACHE, CACHE and more CACHE
NTAP's latest filers provide up to 12 TB of disk space and 6 GB of cache. Symmetrix can scale all the way up to 19.1 TB and 32 GB. That's just the cache. As previously indicated, what makes those large caches effective are the microcodes, the algorithms, the hardwired intelligence that essentially increase the probability that the requested data is read from extemely fast semiconductor memory instead of being read from the mechanical disk arrays. On the write side, a well-designed cache also allows a lot of data to be written to be aggregated in extremely large cache memory prior to being written to the disk drives in a more efficient manner.
EMC is widely considered to have the best microcodes in the storage business so in terms of cache size and microcode technology, NTAP is years behind in this area. The quality of those proprietary microcodes also affect the type of embedded intelligence, embedded diagnostics and storage network management software that can be added to the system.
According to many industry observers, the last vendor to approach EMC in terms of hardware functionality was Storagetek in 1996/1997 with the Iceberg disk array. The rich functionality of the Iceberg and the numerous delays in the Shark were the the main reasons that IBM decided to forge an alliance with arch-rival Storagetek and resell the Iceberg. The bitter falling out between IBM and Storagetek less than 2 years after their alliance is testament to EMC's ability to counter a major threat with simply great hardware and software engineering.
3) LAN-dependency or the lack thereof:
EMC employs a network of networks approach to network storage so they support all kinds of networking protocols. By definition, NTAP's filers are NAS (network attached storage), which are heavily LAN-dependent and one-dimensional . Whatever efficiency they gain from a microkernel-based OS like WAFL, they more than give right back with the inefficiency of a LAN-dependent solution.
As a network interface, Ethernet is 25+ year old. It is cheap and ubiquitous, but it is widely considered to be increasingly structurally inefficient. Machine-to-machine traffic consistently accounts for 30% to 40% of ethernet wire speed and that is expected to trend up with the proliferation of rack-mounted server appliances, the advent of VOIP (voice over IP) and other rich media applications contending with regular users for file-sharing and other applications. Already, some studies are showing that unpredictable network data flows are effectively reducing usable network capacity to an abysmal sub-30% of wire speed. That is great if you're interested in selling storage systems like routers and your customers do not mind having to pay the much higher costs of managing distributed storage. Studies have consistently shown that NAS is only slightly less expensive than disk-attached storage to manage over a congestible network. That doesn't figure to change with high-speed gigabit ethernet and that is why some people believe that Infiniband, with its complex any-to-any connectivity, is ultimately disruptive to the Ethernet interface in the long run.
As things currently stand, Fibre Channel and Gigabit Ethernet are highly complementary technologies with Fibre Channel gaining momentum in SANs and Gigabit Ethernet remaining entrenched in the LAN/WAN. Once Infiniband stabilizes during the roll-outs of 64-bit Unix (Sun, HWP, CPQ, IBM, etc) and 64-bit Itanium and once it starts volume shipment in 2-3 years, it is expected to take clustered servers to another level of functionality and utility. Fibre Channel-based SANs and Infiniband clustered servers are expected to accelerate the development of high-performance and heavily ROI-oriented applications that will drive the virtualization of the LAN/WAN. That creates several inflection points for the server companies and the data networking companies.
So aside from engineering mirroring capabilities into its bigger boxes, NTAP has to develop the EMC-like ability to support a network of networks solutions. That's tough considering that EMC is already leading the way in architecting the second generation storage network with the Fibre Channel interconnect after it totally dominated the original storage networking environment (ESCON), which was invented by IBM for its S390 mainframe. Fibre Channel, by the way, is a lower level transport mechanism that supports ESCON/FICON and IP (Internet Protocol).
As a side bar, proprietary ESCON went from being the only optical interconnect technology between IBM-compatible mainframes and peripherals (file servers, print servers, mail servers, storage systems, robotic libraries, LAN/WAN) gateway servers) to being a mainframe-to-mainframe and mainframe storage-to-mainframe storage interconnect. Fibre Channel will probably follow the same corporate usage patterns in the event that the Infiniband standard should be slowed down by vendor bickering.
Lastly, a file system is the backbone of any operating system. Naturally, NTAP's WAFL Operating System continues to generate a considerable amount of excitement because of the potential to be the de facto network storage standard, but the RAID market has a very poor history of vendors adopting a common standard in anything. The storage business is very unfriendly to standards. Witness the failure of Veritas' decade-long quest to develop a common file system for all storage systems or the fate of Encore's data-sharing technology. Sun acquired Encore in 1997 to spearhead its ill-fated A7000, its last EMC-killer, but it couldn't get Encore's technology to work well with the Solaris OS and it simply didn't have the microcode talent to combat EMC with the big box portion of its big box-small box pincer (tweezer, snoozer) strategy.
I think it's revealing that EMC, the fastest growing storage management software company, with a current annual run rate of $1.4 billion is more interested in working with the key operating system vendors and application vendors to develop a sophisticated SAN file system to increase the functionality and utility of the second generation SANs. I think that's due to the fact that there is a growing demand demand from large customers for ways to reduce the unmanageable inflation of perpetually warring technology abstractions in large scale computing environments. The EMC approach is demonstrably superior, not the least of which is due to the fact that it has the support of many powerful IT organizations in its campaign to make the most popular operating systems work more seamlessly with key applications like relational databases, which in turn, form the foundation of many custom applications that are key to many organizations' strategy of using information as a competitive advantage. Against that backdrop, it's hard to see any new file system blossom into an operating system that will become the de facto standard in the enterprise storage market.
NTAP will probably continue to do well selling WAFL as part of the disk arrays it sells. Because not all disk arrays are created equally or have the same intelligence, I'm extemely skeptical that other RAID vendors will adopt it as a standard. At some point, it is going to be in the interests of IBM (mainframes) Microsoft (Windows), and the strongest Unix vendors (Sun, IBM, HWP, Compaq, Linux) to develop the right network file system for the right applications, probably under a paradigm where clustered servers and networked intelligent RAID arrays with massive processor arrays inside work together to enable more concurrent processing and manipulation of VLDBs (Very Large Databases) in real time and near-real time systems. NTAP may have the chance to license its WAFL technology to Microsoft or the Unix powerhouses in order to displace either the NFS or CIFS file systems, but I view that as a long shot. It may even want to consider working the Linux community on its GFS (global file system).
Here's a useful link to a Linux initiative sponsored by EMC, Seagate and Brocade. EMC completed its Linux program last year and is just waiting for its customers to start using it in mission-critical data centers. Up to this point, customer demand remains tepid.
The Global File System (GFS) is a shared disk cluster file system for Linux. GFS supports journaling and recovery from client failures. GFS cluster nodes physically share the same storage by means of Fibre Channel or shared SCSI devices. The file system appears to be local on each node and GFS synchronizes file access across the cluster. GFS is fully symmetric, that is, all nodes are equal and there is no server which may be a bottleneck or single point of failure. GFS uses read and write caching while maintaining full UNIX file system semantics.
globalfilesystem.org
Since I mentioned Linux, it's also useful to reference the violent history of Unix.....
unix-systems.org
......to give one an idea of the kind of mind-bending industry politics that await nifty file systems that want to be the de facto operating system of............anything, really. |
