Meta Group notes on the IBM Shark
The Bottom line is:
Use it for a competitive hammer.
Product will not be ready for prime time until the end of 2000.
Light on cache, does not yet support Raid-1, has a slow PCI bus, does not yet support Ficon, Fibre Channel, FlashCopy, or PPRC.
Bob Gauthier
In late July 1999, IBM introduced its road map for the long-awaited "Shark"
Enterprise Storage Server (ESS), its best hope for recapturing significant
share of the exploding market for heterogeneous enterprise storage (growing
70% annually through 2004). Much of the ESS hardware foundation will be
delivered in 4Q99 (e.g., hardware, ESCON and Ultra SCSI connectivity, S/390,
Unix, NT, AS/400 attach - see Figure 1 in Addendum). However, users must wait
until 1Q00 and beyond for the more interesting software-driven functionality
(Layers 2-5 of META Group's Enterprise Storage Architecture Blueprint - see
EDCS Delta 846, 18 Aug 1999), which promises a more competitive platform.
While Shark brings IBM into solid hardware competition for scalability (up to
11TB) and performance, users must wait until 1Q00 for FICON (Fibre Connection)
and FC/AL (Fibre Channel-Arbitrated Loop) connectivity and new FlashCopy
(IBM's volume-level SnapShot) and Peer-to-Peer Remote Copy. Moreover, its
virtual implementation will be a homegrown RAID-5 log-structured file
architecture not available until mid-2000.
Indeed, we believe Shark will need a 12- to 18-month gestation period before
gaining the demonstrated credibility required for pervasive deployment in a
global enterprise storage architecture. Mapping IBM's ESS road map against
META Group's Enterprise Storage Blueprint, our analysis indicates that the ESS
will be (by mid-2000) strongest in the two physical component layers and many
elements of the first functional component layer (Layer 2's shared services:
remote mirroring, point-in-time copy, data movement, replication, protection,
etc.). However, we believe IBM will have difficulty delivering robust
functionality in the critically important top three layers (adapter,
application, and management - Layers 3-5) before 2001.
Longer term (2002-04), as IBM rolls out improved enterprise connectivity,
better heterogeneous management disciplines, and a virtual, truly shared data
architecture, we project the ESS will have modest success in regaining some of
IBM's lost market share. Specifically, of a projected 140PBs to be shipped in
2004, we project IBM will account for about 18%, with EMC still dominant at
60%.
Near term (2H99), we believe Shark will have only a minimal effect on global
market share (we project IBM will end 1999 with 12% of the 10,000 high-end TBs
shipped worldwide versus EMC's 63% and Hitachi/HP's 20% - see EDCS Delta 814,
20 April 1999). While this will be due in part to the effect of the global
4Q99/1H00 Y2K procurement freeze, the unavailability of much of Shark's more
interesting software functionality until 1H00 will drive users to a
wait-and-see approach.
On the pricing front, IBM is hungry to hit the ground running and does not
intend to lose (to EMC) on price alone. Specifically, initial bids for Shark
have ranged from free trials (30-90 days with no procurement "strings"
attached) to $0.25-$0.40/MB usable (versus EMC's $0.50+). We believe the ESS
can be an effective tactical (4Q99) competitive threat for large users in
search of leverage against an increasingly dominant (and therefore costly)
EMC. Longer term (2002-04), we believe IBM's ESS will sell for the traditional
PCM discount of 15%+ versus EMC.
The Shark Needs Some Work
We believe Shark's initial iteration will require quick enhancements in
several areas to qualify as a world-class enterprise storage solution:
- RAID-1 mirroring: We believe a significant ESS shortcoming will be its lack
of a RAID-1 mirroring option, which IBM will fix by YE00. Our research
indicates more than 90% of mission-critical disk subsystems are currently
mirrored for protection and performance. As HDA (head-disk assembly)
capacities grow exponentially (IBM will double the current 36GB drive by
YE00), we question the ability of Shark's striped RAID-5 architecture to
deliver the performance required in high-end, mission-critical environments.
Indeed, we project that by YE00 IBM will be forced to introduce RAID-1
mirrored ESS capacity to satisfy performance requirements of users in both
36GB and 72GB configurations.
- Lite on Cache: Shark's 6GB/11TB maximum cache/storage capacity ratio appears
light for the many less than cache-friendly workloads users often encounter in
mission-critical environments. We urge caution in putting large capacity
(>8TB) ESS configurations into high-volume transaction environments until IBM
either demonstrates solid performance (4Q99 scheduled availability of initial
performance data) or increases the maximum cache size (to 12GB in mid-2000).
- PCI Bus: One of Tarpon's (predecessor to the ESS) primary performance
bottlenecks was its relatively slow (132MB/sec) PCI bus, which has been
carried over to Shark. Although the ESS's multiple PCI bus architecture has a
theoretical 800 MB/sec internal bandwidth, we still believe its overhead poses
a potentially serious performance issue. We project IBM will largely eliminate
the problem with the enhanced PCI-E bus of its next (4Q00-1Q01) RS/6000
generation.
________________________________________________________________________
Addendum
Figure 1 - Shark Detail
Hardware - Two 4-way SMP RS/6000 processors, with standard 132MB/sec PCI bus
- Usable capacity: 420GB-11.4TB - Cache capacity: - 3GB per frame = 6GB total
(potentially a performance bottleneck in high-capacity, high-transaction
volume environments) - Nonvolatile storage (NVS) write cache: 192MB per frame
= 384GB total - SSA drives: 9GB and 18GB @ 10K rpm; 36GB @ 7.2K rpm (also
supports most 7133-20 & -D40 drawers) - RAID-5 only: - Base frame: Striped
6+Parity+Spare - Expansion frame: Striped 7+Parity - Connectivity: - Sept.
1999: S/390, Unix, NT, AS/400 - Sept. 1999: ESCON, Ultra SCSI (4 to 32
intermixed server connection ports) - 1Q00: FICON, FC/AL (S/390 only) - 2Q00:
Native Switched Fibre Fabric (Switched Fabric via OEM [Brocade] switch: 4Q99)
- Performance: - Up to 30K cached I/Os/sec - 350 MB/sec top throughput (185
MB/sec sequential throughput) - 800MB/sec usable internal bus bandwidth -
16MB/sec maximum ESCON bandwidth/channel Software - S/390-exclusive
Performance Accelerators (1Q00 availability) Parallel Access Volumes (PAVs)
and Multiple Allegiance - Enables multiple, concurrent I/Os to the same
(logical) volume via UCB alias, significantly reducing pend times; works
dynamically with Workload Manager (WLM) across multiple systems within a
Sysplex. I/O Priority Queuing - Enables ESS or user to prioritize I/O queuing
(can work with WLM in Goal Mode), reducing "device busy" status and improving
performance balance. - S/390 and Open: FlashCopy (1Q00 availability) -
Similar to StorageTek's volume-level SnapShot copy, but not (initially)
implemented via log-structured files (we expect a log-structured deployment in
2H00). An "instant" (two-second) point-in-time copy available for backup, data
mining, etc. - S/390 and Open: Peer-to-Peer Remote Copy (PPRC - 1Q00
availability) - Extension of IBM's volume-level PPRC synchronous copy via
ESCON links between ESS systems. With less than 5% as many PPRC or XRC
installations worldwide as EMC's SRDF, IBM has a significant challenge in
proving their competitive worth. - S/390 only: Extended Remote Copy (XRC
-1Q00 availability) - Equivalent to IBM's old XRC; open systems access only
via XPE copy to 3390-3 type volumes under OS/390.
|
