分类： 服务器与存储
2008-03-10 14:04:20
Readers are the joy of this column. They send observations based on their experience and sometimes attach interesting missives to make their points. Following a recent column on EMC’s DMX3 offering, the mailbag contained a number of criticisms decrying the column's opinions as “stilted” and “biased,” as well as plaudits extolling our “objectivity” and “forthrightness.”
One comment that we took very seriously was a request for a more granular comparison between the two flagship products offered by EMC and Hitachi Data Systems (HDS) that might help consumers understand the nuances of the difference between the two platforms. Shortly after receiving this request, a reader forwarded a slide deck from EMC that his sales rep had used to explain why his product was the better buy over the competitor’s (HDS’) wares.
Disclaimer: We would not have discussed any of the material contained in this deck, labeled “confidential and proprietary,” had it not been for similar claims and comments made in and around an EMC presentation at a storage event we attended a few weeks ago in the Netherlands.
Once the information was shared publicly, we were able to garner a response from HDS Chief Technology Officer Hu Yoshida. Comfortable that the EMC bullet points were “in the wild,” he was willing to offer a response.
Claims and Counterclaims
EMC’s sales tool stated that DMX3 was superior to HDS TagmaStore based on its “higher bandwidth,” “performance that scales with capacity,” “larger mirrored memory,” and “higher backend capacity (more drives).”
EMC DMX3, the materials said, sported 128 GB/s data bandwidth and 6.4 GB/s control bandwidth, which surpassed TagmaStore’s paltry 64 GB/s data and 7.1 GB/s control bandwidth. This kind of bandwidth capacity would help companies that needed to aggregate a lot of data in one box and wanted to avoid the problem of access density that we covered in a previous column.
Yoshida responded, “It is difficult to see how EMC can differentiate data and control bandwidth when they use the same cache and cache paths for both. The TagmaStore Universal Storage Platform (rebranded as HP’s XP12000 and Sun’s SE9900) is the only storage platform that separates control from data into their own cache modules with separate cache paths to each of its 128 processors. The Universal Storage Platform (USP) delivers 68GB/s of data-cache performance and 10GB/s control-cache performance.”
He went on to say that EMC’s 128GB/s Data Bandwidth claim is “equivalent to adding up all the wires in the house and claiming you have 27,000 volts of electricity.” Said Yoshida, “[The] 128GB/s of 'data bandwidth' quickly becomes 16GB/s of cached bandwidth, as there are only 32 bi-directional paths to cache, and only half the bandwidth can be used in any direction at one time. Paths to cache is the limiting performance factor in any high-end, cache-centric disk array.”
EMC’s materials emphasized their claim that DMX3 offers superior performance scaling by noting that “DMX3 delivers scaling performance up to the industry’s highest drive counts,” while “TagmaStore stops scaling at about 400 drives.”
Yoshida’s response was interesting. Polite to a fault, he stopped short of saying that EMC was prevaricating on this point and instead noted, “The TagmaStore USP and its predecessor product, the Lightning 9900V/XP11000, have many installations with over 1000 drives installed. Now with [TagmaStore’s] ability to attach external storage systems, that can increase to many more drives, the claim that this vendor’s products scale up to the industry’s highest drive count does not compute since they have not delivered more than 960 drives to date and have no ability to do external attach of additional storage systems. In addition, most of the DMX’s are installed with RAID 1, which reduces the number of usable drives by half. All other storage vendors have RAID 5 capability that scales.”
Especially revealing was Yoshida’s broader commentary on the differences in storage system design philosophy between the two companies. “EMC believes the future is to build bigger and bigger monolithic storage systems based on their 20-plus year old Symmetric cache architecture. They have failed to recognize that Hitachi has changed the playing field by separating the high function storage controller from the commodity disk array, enabling customers to buy lower-cost tier-2 or tier-3 disks while benefiting from a tier-1 controller. External storage can be brought in and sized to application requirements, taking advantage of the latest technology and historical price erosion of disk. How long will it take to migrate 2000 disks from a monolithic storage array to the next generation of storage?”
Back to the comparison: EMC noted that DMX3 had a larger mirrored memory at 256 GB than did TagmaStore at 128 GB. Quizzically, they added that DMX3 offered up to 204 GB for writes as opposed to HDS with its 44.8 GB. We weren’t the only ones confused by this distinction, and Yoshida contributed what he could to clarify it.
He began by repeating an observation that we had made in our earlier column, “Up until the last release of DMX, EMC has been telling customers that there is no need for mirrored cache memory. What has changed? If it has no value, why does EMC charge extra for this feature?”
He explained the point they were trying to make—and more importantly the point they were seeking to obfuscate—with their emphasis on caching memory size: bigger isn’t necessarily better, he argued. Like so many things, it is how efficiently the memory operates that makes all the difference.
“EMC has a static-cache architecture,” Yoshida explained. “Configurations must be mapped into cache with BIN files. The 32 paths to cache are statically assigned to separate cache segments. An unused cache segment cannot be used by a path that is not connected to it. Control data and data share the same cache and contend with each other for bandwidth. Business Continuance Volumes and SRDF ports must be statically assigned.”
By contrast, he noted, the TagmaStore cache is dynamically configured, with separate accesses to control cache and data cache. The data cache is accessed over a HiStar switch so that all available cache is available to any path. Shadow volumes and replication ports can be assigned as required. Only writes are dynamically mirrored in data cache. Since copies of reads are on disk or external storage, reads do not have to be mirrored. All this provides much more efficient use of cache.”
As an aside, Yoshida confessed some confusion with EMC’s “math” when it discussed its 204 GB of memory “dedicated to writes.”
Said Yoshida, “I am having trouble with this math. If the writes are mirrored, than this means that 102 GBs are write protected. Why the odd number of 204 GB when the customer buys 256 GB? Is the difference, 52GB, reserved for control data? TagmaStore USP can set a limit for the amount of cache that can be used for write mirroring. This can vary. However, if that amount is not being fully used for write protection, it can be used for reads.”
Cache size and operational efficiency becomes important, again, when one considers the balance that must be struck between cache memory and back-end storage capacity to obtain performance gains despite access density increases. Repeating part of its earlier argument, EMC’s sales materials emphasized that DMX3 supported “more than 2,000 drives (960 at general availability), versus 1,152 drives. This, they said, “Supports massive consolidation and 'in-the-box' tiered storage with superior replication functionality.”
Yoshida’s response was succinct, “EMC has yet to deliver 2000 drives. Since their configurations are predominantly RAID 1, the effective capacity will be 1000 drives when they do deliver it.”
Again he waxed philosophical, “HDS believes the days of larger and larger monolithic storage arrays are coming to an end, and the future will be intelligent storage controllers with externally attached disk arrays that can be dynamically provisioned on a granular basis to match application requirements, exploit technology refreshes, and leverage the price erosion of disks. The DMX, with its static cache architecture and BIN file updates, cannot dynamically provision storage. Therefore, it is usually sold with maximum capacity up front, at the then-current disk prices. When it comes time to refresh a DMX with 2000 drives, the down time will be very disruptive, unless it is attached behind a TagmaStore USP.”
The last comment was a zing that wasn’t missed by this columnist. EMC had made additional claims about “HDS’ flawed virtualization” calling TagmaStore a “proprietary solution that nullifies the native functionality of the attached storage arrays.” The deck noted EMC’s superior visibility into—and management of—attached storage, its allowance for “native replication features of externally managed arrays,” its superior “out-of-band” virtualization of storage (with Invista) versus HDS’ “underpowered in-band implementation [of virtualization via the TagmaStore cross-bar switch that] introduces a performance bottleneck and undesirable latency,” and its greater simplicity of operation when compared to TagmaStore’s “operationally complex” solution that requires a lot of manual intervention and mapping definitions.
EMC’s deck went on to tout its own “auto-provisioning capability,” contending that HDS TagmaStore “segments the array into 32 separately managed entities that do not share resources.” The reality of the situation, concluded EMC, was that “HDS segmentation defeats [storage] consolidation benefits—[TagmaStore] relies on human knowledge, responsiveness, and implementation to allocate and adjust over time.” To EMC’s way of thinking, TagmaStore’s “insufficient USP cache and segmentation inefficiencies result in poor utilization and overall decreased performance.” The competitor goes on to argue that its multi-tier storage-in-a-box has “significant total cost of ownership advantages over Tagma.”
Yoshida remained calm in the face of this series of claims, then proceeded to pick them apart. “The HDS virtualization solution is in the market today, and installed by many customers. Many of the 'flaws' that are cited in the marketing materials, such as, 'nullifies native functionality of external attached storage' and 'use of native replication technology' are considered to be strengths by our users because [they enable customers] to use lower-cost storage without sacrificing the advantages of a tier-1 storage controller. EMC has yet to deliver their virtualization solution and appear to be down playing its importance in recent interviews. Ultimately, the market place will decide which solution delivers the most value.”
As for “nullifying the native functionality of attached storage,” Yoshida explained that, from his perspective, this was “goodness.”
“The value of TagmaStore USP (and recently announced NSC) is to provide one common management and one common control-unit functionality for heterogeneous, externally attached storage. It eliminates the need to buy multiple, expensive, high-function storage arrays with expensive software like SRDF. It enables lower cost tier-2 and tier-3 storage to have the same functionality as tier-1 TagmaStore USP, and extends the useful life of older technology storage.”
Yoshida went on to note that TagmaStore’s Universal Virtualization Manager provided excellent visibility into all storage by monitoring and reporting errors, and that additional software such as HiCommand Storage Services Manager provided comprehensive monitoring of system health through standard CIM and SNMP interfaces.
For each additional claim by the competitor, he politely asserted that the author lacked an understanding of TagmaStore’s feature set and presented a set of evidence to back his claims. In the end, to him, it came down to an architectural choice between monolithic, all-in-one-box storage versus a more heterogeneous configuration made manageable by a common switching head and virtualization solution. Yes, he conceded, TagmaStore does usurp certain controller and software functions on the arrays attached to it. However, in so doing, he believed, it created a more manageable infrastructure solution that would enable more choices for data movement and migration, better correlation between storage resources and application requirements, and better technology investment protection and quality of service guarantees.
He chuckled at the statement in the sales materials regarding DMX3’s “auto-provisioning capabilities,” asking the question, “How does one describe auto-provisioning when configurations have to be mapped to cache with BIN files?”
In the final analysis, he offered, “The best way to settle the debate as to whether one product is better than another is to submit both products to a third party for a public benchmark. We invite EMC to submit their latest DMX/Invista to a bench mark with the TagmaStore USP. Jon, would you like to host this in your lab?”
Exciting though the prospect would seem, we have doubts, based on past experience, that participation in such a competitive bake-off would be forthcoming. EMC has not demonstrated great willingness to subject its products to competitive testing when the gauntlet was thrown down by companies such as Softek. So, the real feedback will probably need to come from the market.
It is perhaps anachronistic to make such detailed technical differentiations between array products at a time when most sales are made to the “Front Office,” where business executives of the Global 2000 company reside, rather than the Back Office, where IT lives. But we believe that reporting this debate helps to answer the request of the reader who said such comparisons would make for a more useful column.
