HDS USP V/HP XP24000 is Big on Hype, Falls Short on Real New Functionality

After 32 months, Hitachi Data Systems (HDS) updated their high-end TagmaStore USP on May 14. The modest improvements of the new USP V will not significantly change the competitive landscape. HDS offered little new functionality and failed to match all of the rich features EMC delivered in Q1 with Enginuity 5772. HP co-announced this subsystem as the XP24000 on the same day. Sun has yet to announce their version of this product.
Highlights of the announcement include:
•        Critical Areas: The announcement contains no improvements in Internal Capacity, Security, QoS, and Upgrade Path. Hitachi emphatically declared the end of disruptive upgrades when TagmaStore was first introduced. The reality is that Hitachi requires a disruptive forklift from the current TagmaStore USP.
•        Performance: Hitachi made a number of disjointed performance claims, the major one being 3.5 million IOPS. This 3.5 million number is attainable by customers only not using global cache or disk storage, which is completely unrealistic in any customer environment. We address Hitachi’s performance claims below.
•        Virtualization: In a sign of confusion, HDS claims up to 500% and HP up to 250% improvement in virtualization port performance. One other unrealistic claim is that USP V can manage 247 petabytes externally—up from the likewise unrealistic 32 PB with TagmaStore. There are major limitations with Hitachi’s virtualization approach and its deployment. There are links below to two competitive articles that address this issue.
•        Hitachi Dynamic Provisioning: This new feature provides thin provisioning from a virtual pool of storage. No technical details or GA date were provided, but Hitachi did admit to three significant limitations of this offering:
        Cannot thin-provision remote replicated volumes.
        Cannot thin-provision virtualized volumes. Thin provisioning technology is most applicable to Tier II and III data. However, this is precisely the data that cannot be thin-provisioned in a Hitachi virtualized environment.
        Cannot deploy on N-1 generation Hitachi arrays.
•        Green Storage: Hitachi claimed up to a 50% energy efficiency improvement over the TagmaStore USP but offered no substantiated evidence. Hitachi does not provide a power calculator (as EMC has for more than a year) so customers don’t know how much better or worse the USP V is for specific configurations.
•        Large Logical Storage Pools: This is the ability to stripe data across dozens or hundreds of drives, which Hitachi refers to as “wide striping.” EMC has offered striped metavolumes for a decade.
•        Service Oriented Storage Solutions: This is just a repackaging of existing services. They have not changed their offering of services.
•        Pricing: No details provided by Hitachi.
Product Names
The table below summarizes the product names for the different vendors. Sun has not yet made an announcement.
Item        HDS        HP        SUN
Current High-End Array        TagmaStore USP        XP12000        Sun StorageTek 9990
New High-End Array        USP V        XP24000        Not Announced
Thin Provisioning Software        Dynamic Provisioning        Thin Provisioning        Not Announced
Virtualization Software        Universal Volume Manager (UVM)        External Storage XP        Universal Volume Manager (UVM)
What Was Announced
The USP V offers the following specifications:
Category        Current TagmaStore USP / XP12000        New USP V / XP24000
Architecture        Internal switches: 4         Internal switches: 4
        128 × 400 MHz MIPS processors        128 x 800 MHz MIPS processors
        Maximum Cache: 256 GB        No change
        Maximum Shared Memory: 12 GB        Maximum Shared memory: HDS 32 GB / HP 24 GB
Connectivity        Fibre Channel = 192 ports
FICON = 96
ESCON = 96
Gigabit Ethernet for replication = 0
iSCSI = 32
NAS Blades = Up to 8 blades        Fibre Channel = 224 ports
FICON = 112
ESCON = 112
Gigabit Ethernet for replication = 0
iSCSI = dropped this capability
NAS Blades = dropped this capability
        Addressing for up to 65,536 mainframe or 16,384 Open Systems logical volumes        Addressing for up 65,536 (mainframe or Open Systems) logical volumes
        4 Gb/s front-end FC and FICON (host attachment)        No change
        64 × 2 Gb/s FC-AL (drive attachment)        64 x 4 Gb/s Switched FC (drive attachment)
Performance        Claim 2.5 million IOPs when confined to “special front end” workloads.        Claim 3.5 million IOPS when confined to “special front end” workloads.
        Internal cache data bandwidth: 68 GB/s        No change
        Internal shared memory bandwidth: 13 GB/s        Internal shared memory bandwidth: 38 GB/s
Capacity        1,152 maximum drives in up to 5 frames        No change
        Supported drives:
73 GB, 15 K
146 GB, 15K
300 GB, 10 K        No change
        Maximum internal capacity of 1,150 x 300GB = 332TB        No change
        External virtualized capacity of 32 PB; Still not a realistic figure.        External virtualized capacity of 247 PB; An even less realistic figure.
Software Functionality
        Thin Provisioning is not backward compatible with previous generations of storage arrays.         Thin Provisioning of storage—discussed later in this document.

Hitachi Large Logical Storage Pools—used to facilitate Thin Provisioning through “wide striping”
What Hitachi Did Not Announce
Hitachi did not address many areas that are critical to customers:
•        No improvements to previously announced advanced business continuity features like three-site disaster recovery (announced over a year ago but still not delivered), no combined mainframe and Open Systems consistency groups, and no multi-frame consistency groups in Open Systems. The one enhancement was a claim for an improvement of up to 200% (170% according to HP) of TrueCopy synchronous link performance.
•        No drive count increase—still only a maximum of 1152 drives, less than half the 2400 drives of DMX-3.
•        No high-capacity drives—limited to 300 GB while DMX-3 has shipped 500 GB for over a year.
•        No additional global cache—ltill 256 GB versus 512 GB for DMX-3.
•        No new security features—they still don’t offer features as powerful as what EMC announced in 5772.
•        No additional QoS features like Symmetrix Priority Controls and true Dynamic Cache Partitioning.
•        No upgrade path—Hitachi requires a disruptive forklift from the current TagmaStore USP, in either a virtualized or a non-virtualized environment.
•        No additional virtualization features—Hitachi’s virtualization still represents a single point of failure in the SAN. Refer to the links at the end of this document for more information on the limitations.
What Hitachi Did Announce
Performance Increase
Hitachi increased the speed of the internal processors on the front- and back-end directors from 400 MHz to 800 MHz MIPS. This 2X improvement in processor speed does not necessarily translate into a performance increase of the sub-system. It depends on where the bottlenecks are, which tend to be in different places for different workloads.
The announcement included a claim of up to 3.5 million IOPS. This is up from the 2.5 million IOPS claim on the previous TagmaStore. EMC’s Technical Competitive Analysis Group proved that the 2.5 million IOPS is only possible for a very special “front-end read hit” type of workload where the I/O is serviced from a special buffer in the front-end director, and not from the global cache. This type of an I/O is completely unrealistic in any customer workload environment.
HP specifications indicate that the Read-Miss performance of the XP24000 has only increased by 10% over the XP12000. Even with this increase, DMX-3 is superior to the XP24000 in a Read-Miss environment.
Customers are more concerned about real-world application workloads, and it is in the real world that DMX-3 continues to lead the way. DMX-3 has much better I/O service times than the TagmaStore USP has across workloads, and especially better in consolidated workload environments. Your local Competitive Guru has access to a wealth of more detailed performance information.
Hitachi’s claim for higher internal bandwidth is limited to an increase in speeds (from 13 GB/s to 38 GB/s) to and from the Control Memory. Adding this increase to the existing “data” bandwidth of 68 GB /s (no change from TagmaStore) results in a total of 106 GB/s. A similar calculation of “internal bandwidth” for the DMX-3 results in 134.4 GB/s (128 for data, 6.4 for messaging).
Since Control Memory and Control Memory bandwidth do not directly affect the movement of customer data, but rather deal exclusively with internal system functions, there is no obvious customer value associated with this hardware change. The “increase” in total system bandwidth is therefore extremely misleading. There has been no increase in bandwidth as it relates to customer data.
4 Gb/s Front and Back-End
With this announcement, Hitachi is announcing 4 Gb/s Fibre Channel back-end directors. This capability is expected on all new arrays to be introduced in 2007. As Joe Tucci mentioned in our Q1 earnings announcement, EMC is currently shipping 4 Gb/s front-end directors and expects to be shipping 4 Gb/s back-end directors in about the same timeframe as our competition. It should be noted that 4 Gb/s back-end directors provide minimal or no performance gain for most “cache friendly” workloads.
Maximum Connectivity up to 224 FC, 112 ESCON, and 112 FICON ports
Hitachi announced connectivity increases of up to any allowed combination of 224 FC, 112 FICON, and 112 ESCON ports. HDS arrays have offered more ports than DMX-3 for some time. EMC recommends the use of SAN switches and directors for consolidation of storage connectivity rather than increasing the number of front-end connections to the array. Consistent with server consolidation and storage consideration, SAN path consolidation presents a compelling Total Cost of Ownership (TCO) value proposition. Not only are SAN ports less expensive than array-based ports, but SAN switch ports enable customers to have flexible SAN paths that are easily modified or repurposed.
The facts are that SAN switches lower TCO, reduce complexity, and save power while enabling array ports to run at a higher utilization. Note that the main reason HDS needs more FC ports is that when virtualizing external arrays, many of the front-end ports are used to connect to the external arrays being virtualized. In fact, the way Hitachi gets to these high port counts is by decreasing the number of ports available for internal storage, thus reducing the amount of capacity that can reside on the USP.
Virtualization
Storage virtualization took a prominent place in the announcement.
•        Support for up to 247 PB of external storage, up from 32 PB in the previous USP generation (we haven’t heard Hitachi offer any additional benefits to existing TagmaStore customers). These numbers are purely theoretical and unrealistic in a practical sense.
For example: A single new USP V offers 256 GB of cache to support up to 247 PB. This associates 1,000 TB of capacity per 1 GB of cache—a completely unrealistic ratio of disk capacity to cache size. A virtualized environment of 247 PB is unrealistic because it would require volumes greater than 4TB.
•        Performance improvement for a virtualized port. In a sign of confusion, HDS reports up to 500% improvement, HP reports up to 250%. TagmaStore technical manuals explicitly state that OLTP applications, or essentially any data that requires a high level of performance, should not be deployed on virtualized disk. Only Backup and Archival applications are recommended for virtualized data. At this time, Hitachi has not yet recommended which applications should be deployed in a virtualized environment with the USP V.
It is important to understand that HDS positions their virtualization as the future of “intelligent storage,” not requiring customers to replace older arrays, nor migrate large amounts of TB when replacing them. These are false and misleading statements as they do not address the need to migrate data when the USP array or the virtualized array needs to be replaced.
Hitachi Dynamic Provisioning (Thin Provisioning)
No GA date or technical specifications were released about Hitachi Dynamic Provisioning. HDS and HP did say that their implementation will not work with remotely replicated or virtualized storage volumes—only volumes internal to the USP V—nor will it work on N-1 generation Hitachi arrays. We cannot comment on specific details until this feature is made available and we have a chance to analyze it. Right now, it is just Hitachi marketing hype.
Readers should be aware that no other high-end storage vendor offers this capability. It will be critical to understand HDS’s specific implementation in terms of ease of use, performance impact, and reliability.
For example, during the announcement, HDS clarified that their implementation has soft and hard alerts when the storage pool capacity is being exhausted, but no word on automatic re-allocation of storage. If this is correct, applications reaching that usage level could be in danger of crashing should the customer miss the alerts. On the announcement webcast, HDS CTO Hu Yoshida was asked what would happen if the alerts were not acted on and the array ran out of physical storage. Hu replied “bad things” and it would require manual reconfiguration to fix.
Hitachi has struggled with new software functionality in the past—witness their first implementation of Hitachi Universal Replicator, which was fraught with problems and delays—so we expect their first implementation of thin provisioning to be problematic as well.
‘Green Storage’
Hitachi seems to now understand the importance of energy efficiency in data centers. During their announcement, they claimed up to a 50% improvement in power and cooling over TagmaStore. Once again, these claims are not realistic, as they compare the real capacity they would have to allocate on the older array to match the capacity presented by Hitachi Dynamic Provisioning on the new array. When power and cooling numbers are compared under similar configurations, the new USP V spec sheet shows numbers that are about 10% less than the TagmaStore but no justification is given.
Note that DMX-3 has offered customers significant power and cooling savings over the TagmaStore arrays. The EMC power calculator accurately provides actual power draw of ANY DMX-3 configuration.
While the EMC power calculator has been in existence for more than a year and has been extremely popular, Hitachi has yet to follow suit.

;P  ;P ;P ;P

太长了，能否翻译再简短些？呵呵

够乱的！