IP Storage Hot Topics
Here's what's hot in the IP Storage world...
What is the difference Between an HBA (Host Bus Adapter) and a Dynamic TOE (TCP/IP Offload Engine)?
HBAs
and TOE accelerators typically both occupy a PCI slot on the server.
When the host initiates a read or write, the HBA acting as the storage
interface, receives the request. The HBA hosts several on-board
functions that enable it to move data. This includes the iSCSI
initiator (sometimes called a hardware initiator as it is resident and
executed in silicon on the HBA), its own TCP network stack (as opposed
to the host OS), and local memory and processing. Combined, these
functions free up the host processor and memory subsystems and
facilitate data placement. This can provide benefits in preserving CPU
and memory cycles and generally allow the server to perform better. How
the components, be they software or hardware, are integrated varies
from implementation to implementation. Because the host is fully
offloaded from the burden of processing and moving iSCSI traffic, the
implementation is often referred to as Full Offload. Because the iSCSI
HBA hosts its own complete TCP network stack and Initiator, there have
been reported compatibility problems with applications executed on the
host. Users evaluating iSCSI HBAs should check with the host OS
provider to ensure compatibility and support. Two network stacks, one
on the host and on the iSCSI HBA, each capable of setting up and
tearing down connections can lead to problems.
Because iSCSI HBAs
provide full offload support, host CPU utilization can be low and
efficiency high. While host efficiency can be maintained at high levels
across a wide range of traffic patterns (all block sizes/read-write
mixes), the same cannot be said for data throughput. The onboard
processing can limit the iSCSI HBAs ability to achieve sustained line
speed performance, most notable for larger block sizes. So while the
host’s utilization is low, the iSCSI HBA itself can become the
bottleneck limiting performance.
One of the reasons
that customers are attracted to an iSCSI HBA is the simplistic boot
from SAN feature. It is relatively easy to implement for a small number
of servers. However, each server required to boot from the SAN must be
equipped with an HBA, each having a unique boot image. Each HBA has its
own BIOS to support boot. Customers utilizing Dynamic TOE Accelerators
can also boot from the SAN using network boot. While this requires a
DHCP server, the boot process when large amounts of servers are
deployed, be they equipped with Dynamic TOE Accelerators or not,
becomes greatly simplified. From a storage perspective this is
attractive as one boot image can serve multiple servers. A Dynamic TOE
Accelerator in contrast to the iSCSI HBA is architected to be
compatible with the host TCP network stack. The Dynamic element is
executed when the host transfers connection state to the Accelerator
which assumes control of TCP processing and data transfer. Unlike an
iSCSI HBA, a Dynamic TOE Accelerator makes use of custom silicon
optimized as a DMA engine to move data extremely quickly as opposed to
a HBA which uses a general purpose processor.
iSCSI Management API version 1.1.6 is approved as an
ANSI standard
ANSI INCITS 411-2007)
The IMA is a specification for a C language based API
for managing iSCSI capable HBAs and NICs, along with the device drivers
that control IMA provides APIs to perform a variety of different functions
including:
As an "open systems" specification, the IMA is intended
for use across a wide variety of operating systems including Windows®,
Linux®, Solaris™, HP-UX, and AIX®. Storage management applications will
use the IMA to help them manage iSCSI SANs.
What is Fibre Channel over Ethernet (FCoE) ?
Fibre Channel over Ethernet is a proposal for a new protocol specification
which has been submitted to the INCITS T11 Committee. FCoE is a means of
transmitting Fibre Channel commands natively over Ethernet by using an
optional transport mechanism instead of TCP/IP, while maintaining backward
compatibility with existing Fibre Channel endpoint infrastructure. FCoE is on
an aggressive schedule in T11, and there is a good chance that it will emerge
as a new ANSI standard in the 2009 timeframe.
What is the buzz about Green Storage? How does it relate
to IP Storage?
If you have been reading news related to storage, you probably come across
the word "green" quite a few times in the last year or so. So what is the
entire buzz about Green Storage and how it's going to affect IP
Storage?
A new taskforce and a Technical Working Group (TWG) has been set up within
SNIA to provide education on Green Storage, and to create power efficiency
metrics for storage systems. The SNIA Green Storage initiative (GSI) will
collaborate closely with the Green Grid (a vendor consortium focused on power
efficiency at the data center level) to ensure consistency of metrics and
education. Although the GSI is focused primarily on power efficiency, other
'green' topics have not been ruled out.
How does IP Storage plays a role in this effort?
Over the last few years iSCSI matured rapidly with growing acceptance and
credibility among SMBs and Enterprise Customers for cost effective storage
consolidation solutions for mission critical business applications,
particularily for Windows environments. 10 Gigabit Ethernet deployments are
emerging with improvements like scalable networking. Regarding TCP
offloading, we are opening the flood gates for large scale storage
consolidation coupled with ease of use
Consequently, iSCSI is playing, and will continue to play, a major role in
innovations that will make data centers consume less power both directly and
indirectly. The combination of easy to use, high performance arrays based on
the latest drive and power technologies, the widespread implementation of
array-based virtualization, thin provisioning and Data De-Duplication all
indicate that when power efficiency metrics are available for storage
systems, IP Storage will be seen as a strong contributor to 'green' data
centers.
iSCSI's role in the next-generation
datacenter
The 'data center of the future' is increasingly an industry topic of
conversation. Over the coming months this will continue to increase as the
complex challenges around power and space intersect with increasing focus on
a 'unified' data center fabric - with a single network architecture for
storage networks, servers and services. It seems likely, at this point, that
10 Gigabit Ethernet (10GBE) will play an important role here.
10 Gigabit Ethernet is available now for switch, server and storage
interconnects. Remote Direct Memory Access (RDMA) work is well advanced in
the IETF - enabling low-latency sesrver clustering. Latency and performance
enhancements for Ethernet are well under way in the IEEE 802.3 Working Group.
Fibre Channel over Ethernet (FCoE) is very active in ANSI T11 - ensuring a
future seamless transition to a 10 Gigabit Ethernet fabric infrastructure for
today's Fibre Channel HBA's and storage arrays. And, iSCSI is will
established as a mainstream navite storeag protocol for IP-based SAN
solutions.
Customer can confidently buy iSCSI-based SAN solutions today, knowing that
they have a smooth path to 10 Gigabit Ethernet, the various transport-level
latency and performance enhancements, and the unified fabric services we are
likely to see in the next generation data centers.
iSCSI over Infiniband: does this make
sense?
With iSER protocols ratification, and more products coming online, users are
looking for validation that iSCSI and Infiniband are complimentary
technologies ready for primetime. In a nutshell: infiniband offers an
efficient point to point architecture between processors and I/O traffic that
offers increased bandwidth and "switching" functionality. In theory the
simplicity of iSCSI deployment combined with the low-latency of Infiniband
can offer a excellent high-performance storage infrastructure for
applications beyond Infiniband's stronghold in high performance computing.
The iSER protocol allows iSCSI traffic to bypass TCP processing altogether by
using RDMA increasing performance and decreasing overhead. With today's
workloads and processing power the overhead of the TCP protocol isn't
problematic since typical systems are still significantly underutilized, but
this may change in the future as system utilization increases, standard
networking speeds increase beyond 10Gb/s and applications begin to take
advantage of the power of the newer interconnects.
Storage Grids
Grids have been widely discussed in technical circles over the past few years
as the next killer app to take advantage of the relative power of today's
'low-cost' hardware. Storage grids are next on the agenda and are being
utilized in some shape or form by clustered NAS operating systems and
Google's file system. Each of these file-centric technique offers a way to
consolidate disparate storage resources into one giant, redundant pool. The
transition to grids will occur once more storage services happen at a virtual
layer, where physical storage resources are pooled together and services
happen independent of the disk controllers. As end-user begin to demand and
deploy more open-standards, and vendors increase interoperability, block
level storage grids utilizing IP protocols should become more prevalent via
advanced storage virtualization techniques and increased consolidation and
centralization of server, network and storage resources.
Grid, Thin Provisioning and Cloning
Usage of a Grid Approach in combination with thin provisioning and cloning
reduces the amount of individual copies of operating systems disks and
provides a minimum number of "Golden Images" which can be used to serve a
large number of disk-less blade servers -- this greatly diminishes the number
of physical drives needed which also has an impact on an achieving a
"Greener" environment.
