Sie sind auf Seite 1von 3

tech report

It’s best to store remote-office data and host remote-office applications


in the central data center. Various products can eliminate TCP/IP transmission
issues that might hinder those moves. By Marc Staimer

s
48
everal recent surveys by analyst firms document the perception
that the amount of data within remote offices and branch offices (ROBOs) is
greater than the amount of data within the data center. In an era of growing
compliance regulation, and increased internal and external threats to data, man-
aging and protecting ROBO data has become a big problem.
ROBO users require and expect the same level of services and support as users
in the primary data center. This means ROBO data accessibility, data protec-
tion and application response times must be equivalent to those local to the data
center. The requirements are simple. Delivering on those requirements isn’t.

protect
Storage January 2007 Illustration by Neal Aspinosa
remote-office data
Skinny pipe TCP WAN optimization and WAFS product comparison
Riverbed Packeteer Inc. Cisco Systems Inc. Juniper F5 Networks Inc. Citrix Systems Inc. Expand SilverPeak
WAFS support Technology Inc. iShared Wide Area Networks Inc. WANJet NetScaler Application Networks Inc. Systems Inc.
Steelhead Application Sequence Accelerator Accelerators NX Series
appliances Services Reducer/Mirror
pabilities in diminishing the throughput effects of
CIFS chattiness reduction 3 3 3
3 3 3 3 3 packet loss. WAN packet loss is a fact of life. Most
NFS chattiness reduction 3 3 3
3 3 telcos will claim incredibly low packet-loss numbers
E-mail 3 3 3 3 3 of 0% to .001%, but those seem to be the exception,
SQL databases 3 3 not the rule. Users, on the other hand, say their WAN
Static Web 3 3 3 3 WAN accelerator 3 3
3 3 3 3 WAN accelerator 3 3 3
packet-loss rate ranges between .1% to 6%, accord-
Dynamic Web
Deduplication 3 Files only Files3only 3 3 ing to various surveys. Although that may not seem
Compression 3 3 3 3 3 3 3 3 like much, it’s devastating to TCP/IP throughput.
Application visibility 3 Partial 3
Partial 3 3 TCP was created with very short LAN distances in
and management
3 3 3 3 3 3
mind. When packets are lost or dropped, TCP can’t re-
Caching
Quality of service 3 Cisco IOS 3 3 order packets, so it must retransmit every packet be-
Encryption 3 Cisco IOS 3 hind the lost one. This isn’t such a big deal on a LAN
Local storage (hybrid) 3 3 3 3 because the distances are short, acknowledgement
Fat ERP client support Future 3 of packet loss is quick and not much data has to be
retransmitted when packets are lost. That isn’t the
case when traversing a WAN; as distance increases,
Storage managers have two main ways to protect Unfortunately, the vast majority of applications Data Services [WADS], Wide-Area Application roundtrip time increases and more packets fill the
ROBO data: create a mini data center for each ROBO, were designed for local LAN environments. They Services [WAAS] and Wide-Area Data Manage- pipe. This means more packets must be retransmitted
or centralize ROBO applications and data at the pri- don’t take into consideration WAN issues such as ment [WADM]) when a packet is lost.
mary data center. The first option is very expensive, bandwidth limitations, TCP/IP packet latency (TCP 3. Distributed ROBO backup to disk DRO significantly reduces the effects of packet
management is complex, highly qualified IT staffers overhead), speed of light latency (distance) or TCP/IP We’ll look at how each of these three technologies loss by reordering the packets, so it only has to re-
are required at each ROBO and things rarely work packet loss (retransmissions). These WAN issues can resolves WAN issues, strengths and weaknesses, and transmit the lost packets and nothing else. DRO
well. reduce throughput by as much as 95% of the band- where each is ideally suited for deployment. data throughput is often an order of magnitude
Centralizing ROBO applications and data in the width. As bandwidth is reduced, user frustration greater than native throughput. Sometimes it’s ac-
primary data center offers the following potential grows because of agonizingly slow application re- Fat pipe TCP WAN optimization tually greater than the rated bandwidth (depending
advantages: sponse times with reads and writes that are measured Fat pipe TCP WAN optimization (DRO) is designed on the amount of bandwidth). These quantifiable
• Increased IT productivity by leveraging IT ad- in minutes vs. milliseconds. to move large amounts of data in single or multiple throughput improvements can be traced directly to
ministrator assets within the data center Three technologies enable the centralization of streams over “fat pipes” at a speed of between 5Mb/sec the cumulative effects of deduplication, compression,
• Reduced investments in hardware, infrastructure, data by mitigating or eliminating most WAN prob- to 1Gb/sec. DRO minimizes the effect of distance by optimal window sizing, stacking, TCP termination
software, personnel, training and maintenance, lems (with the exception of speed of light latency or using a TCP protocol-enhancing proxy (typically and packet-loss mitigation.
and increased data center asset utilization distance). These include: a derivative of TCP or UDP) that allows
• Better, faster service and simpler management 1. Fat pipe TCP WAN optimization (a.k.a. Data it to do things that can’t be done with AS BANDWIDTH IS REDUCED, USER FRUSTRATION
• Consistent data protection that meets IT and reg- Replication Optimization or DRO) standard TCP/IP v4 or v6. For example, GROWS BECAUSE OF AGONIZINGLY SLOW
ulatory compliance requirements
• Easier sharing of data among ROBOs
2. Skinny pipe TCP WAN optimization, plus Wide-
Area File Systems or WAFS (a.k.a. Wide-Area
the protocol-enhancing proxy allows
DRO to terminate TCP at each end of APPLICATION RESPONSE TIMES WITH
the pipe, which completely eliminates READS AND WRITES THAT ARE MEASURED IN
TCP latency overhead. That, by itself, MINUTES vs. MILLISECONDS.
Data Replication Optimization (DRO) product comparison would greatly increase the effective data
throughput, but there are other DRO tricks that Sometimes DRO throughput can be too efficient.
NetEx Riverbed F5 Networks Inc. Packeteer Inc. Citrix Systems Inc. increase throughput. It can completely seize the entire fat pipe, making
DRO focus HyperIP Technology Inc. WANJet SkyX Accelerator WAN Scaler
Steelhead appliances DRO deduplicates data on a block level, which re- it available only for those applications running
Extensive DRO app quals 3
quires a lot less data to traverse the fat pipe. It can also through the DRO, which can block other apps from
Packet-loss mitigation 3 3 Limited compress the deduplicated data, as well as adjust the accessing the WAN. One way around this is to en-
Rate limiting 3 3 window size and payload to decrease the number of sure the DRO product has rate-limiting capabilities
B/W support to OC3 3 3 3 3 3 roundtrips required to move that data between the pri- or assigning only a fraction of the WAN bandwidth
(155Mb/sec)
mary data center and remote or branch office. (When to the DRO (see “Data Replication Optimization
B/W support to OC12 3 3 3 3
(655Mb/sec) the TCP window size and payloads are increased, it [DRO] product comparison,” p. 50).
End-to-end network visibility 3 3 Limited Limited usually makes the effective data throughput far more DRO is an excellent choice when bandwidth is
Bandwidth optimization 3 3 3 3 3
sensitive to packet loss caused by high bit-error rates, high, packet loss exceeds .1%, distances are greater
Compression 3 3 3 3 3
Deduplication 3 3
network jitter and network congestion.) The net effect than 300 kilometers and when data migration ap-
Encryption 3 3 3 is a greatly increased data throughput rate. plications have to move massive amounts of data
The real throughput magic comes from DRO’s ca- across the WAN. DRO isn’t normally a good choice

50 Storage January 2007 January 2007 Storage 51


Distributed ROBO backup-to-disk product comparison
ASIGRA EMC EVAULT IRON SIGNIANT SYMANTEC
INC. CORP. INC. MOUNTAIN INC. INC. CORP.
Televaulting Avamar InfoStage, LiveVault, Mobilize for Pure
Axion ArcWare, InSync, Remote Data Disk
Replicator Continuum, InControl Protection some combination of deduplication, compression, tems in servers, desktops and laptops that range in
Desktop (RDP)
SERVER (S), DESKTOP (D), LAPTOP (L) SUPPORT sequence caching, TCP and UDP acceleration, band- size from a single user to hundreds of users. To do
Windows S/D/L S/D/L S/D/L S/D/L S S width management, multithreading, quality of serv- that effectively, these backup products or data serv-
Red Hat Linux S/D/L S/D/L S/D/L S/D/L S S ice (QoS) and path optimization. ices need to be more efficient than traditional back-
Novell SUSE Linux S/D/L S S WAFS goes a step further by providing accelerationup or replication technologies, and easier to use.
Novell NetWare S/D/L S/D/L for specific applications such as CAD/CAM, print, Most of the products provide multiple point-in-time
Mac OS X S/D/L S
HP-UX S/D S/D S
Web caching, e-mail, DBMS or enterprise resource data versions similar to continuous data protection
HP Tru64 Unix S/D planning. The combination of skinny pipe WAN (CDP) products.
Sun Solaris S/D S/D S/D S/D S optimization plus WAFS is an outstanding way to The ROBO WAN efficiency comes from dedupli-
IBM AIX S/D S/D S/D centralize ROBO data. One way WAFS radically im- cation (locally and globally), transmission of delta
EMC VMware S/D S/D proves the performance of some of these apps is by changes and compression of the remaining data.
IBM iSeries OS/400 S S
reducing the application protocol’s “chattiness.” Distributed ROBO backup-to-disk data is deduped
MAIL SERVER AND DATABASE SUPPORT
The Common Internet File System (CIFS) is an across the WAN in flight and at rest. DRO, skinny
Microsoft SQL Server 3 3 3 3 3
Microsoft Exchange Server 3 3 3 3 3
excellent example of a very chatty application pro- pipe and WAFS data is only deduped in flight. In ad-
and Outlook 2000/2003 tocol that requires numerous commands for every dition, local deduplication takes place at the ROBO
Oracle 8 and above 3 3 3 3 transaction. Each command creates a roundtrip location, while global deduplication takes place
DB2 3 from the initiator to the target and back, with eachat the central data center site. Global deduplication
MySQL 3 3
PostgreSQL 3 3
roundtrip adding latency (delay) to the transaction.removes the duplicates among all of the ROBO
IBM Lotus Notes/Domino 3 Latency increases exponentially as distance in- sites.
Server creases because of the cumulative effect of all the The significantly reduced protected data will usu-
Novell GroupWise 3 roundtrips. WAFS turns around the CIFS commands ally make disk storage less expensive than backing up
ADVANCED FUNCTIONALITY locally and stacks them so that most or all of the com-
to tape. In fact, there’s no need for tape, tape libraries
Continuous data 3 3 3 mands required in the transaction traverse the pipe or virtual tape. Recoveries to the ROBO are also a
protection
Autonomic verification 3 3 great deal faster since the data comes di-
and healing (data cleansing)
3
CENTRALIZING ROBO APPLICATIONS rectly from disk. For archiving, backups
AND THEIR DATA AT THE PRIMARY DATA CENTER
No host software required can be moved to tape or optical media as
(agent)
3 3 3 the protected data ages.
Local deduplication (SIS)
Global deduplication (GSIS) 3 3 3 HAS ENORMOUS ECONOMIES OF SCALE. Distributed ROBO backup to disk is
Enterprise scalability 3 3 available from the following vendors: Asi-
Bare-metal restore 3 at the same time. ROBO performance increases by gra Inc., EMC Corp./Avamar Technologies, EVault
Archival 3 3 3
ROBO (local) recovery 3 3 3 3 3
“orders of magnitude,” allowing many ROBO appli- Inc., Iron Mountain Inc., Signiant Inc. and Syman-
WAN optimization 3 3 cations and their associated storage to be centralized tec Corp. (see “Distributed ROBO backup-to-disk
(compression, delta and consolidated in the primary data center (see product comparison,” p. 52). Distributed ROBO
changes, other)
Native backup to tape 3
“Skinny pipe TCP WAN optimization and WAFS backup to disk is an excellent choice when data pro-
and tape libraries product comparison,” p. 50). tection is the primary issue or there are numerous
File search capabilities 3 3 3 3 3 Skinny pipe TCP WAN optimization plus WAFS mobile users. It’s not a good choice when the primary
Single pane-of-glass 3 3 3 3 3 is an excellent choice when there are quite a few ROBO issue is performance when working with cen-
management (Web portal)
Secure encryption 3 3 3 In flight 3
ROBOs, the bandwidth to the ROBOs is less than tralized data center applications.
in flight and at rest 5Mb/sec and performance comparable to that of Centralizing ROBO applications and their data at
locally hosted applications for ROBO users is im- the primary data center has enormous economies of
portant. Skinny pipe TCP WAN optimization plus scale. To do so effectively also requires implementing
when the bandwidth to ROBOs is relatively “skin- bandwidth less than 5Mb/sec; typically, they have a WAFS isn’t a good choice when the amount of data technologies that eliminate TCP WAN transmission
ny” (typically 1.5 Mb/sec or less), there’s no quan- bandwidth of 1.5Mb/sec (T1) or less. Skinny pipe moved between locations is very large or for laptop issues. When an organization contemplates how it
tifiable packet loss and the distances are very short. TCP WAN optimization is functionally similar to mobile users. will centralize distributed ROBO data, it must first
DRO also isn’t a good choice if the amount of data DRO; the key difference—and it’s not a trivial one— determine which applications will be centralized and
moved over the WAN is light or insignificant, such is the size of the bandwidth pipe. Algorithms and tech- Distributed ROBO backup to disk the application’s required performance. It’s possible,
as Web applications. niques that work with fat pipes don’t work nearly Distributed ROBO backup to disk is the first data and in many cases highly likely, that multiple tech-
as well with skinny pipes and vice versa. Skinny pipe protection technology designed specifically for nologies should be deployed to meet the organiza-
Skinny pipe TCP WAN optimization plus WAFS TCP WAN optimization has effective data throughput ROBOs. It has been around for nearly 20 years, but tions needs. 2
Skinny pipe TCP WAN optimization and WAFS are results similar to DRO. the technology has only recently been available for
often features and functions of the same WAN opti- Vendor methods vary, with widely differing re- license by end users. Distributed ROBO backup to Marc Staimer (marcstaimer@earthlink.net) is president
mization controller. Skinny pipes are defined as sults depending on the data. In general, vendors use disk protects data for a wide variety of operating sys- of Dragon Slayer Consulting.

52 Storage January 2007 January 2007 Storage 53

Das könnte Ihnen auch gefallen