Simple, Reliable Performance for iSCSI Connectivity

8K4K2K 32K16K 64K 128K 256K 512K 1M1K512
0
200,000
400,000
600,000
800,000
1,000,000
IOPS
I/O Size (Bytes)
Intel® 82599 10 Gigabit Ethernet Controller Single Port Read/Write IOPS Performance
One Million IOPS
Figure 1. iSCSI read/write performance on the Intel® 82599 10 Gigabit Ethernet Controller, expressed in terms of IOPS (I/O operations per second).
1
Platform Advancements
10GbE Intel Ethernet Server Adapters
support data center bridging (DCB) in
hardware to provide the lossless or no-
drop Ethernet fabric needed for storage
trafc. This set of standards enables
better trafc prioritization over a single
interface as well as an advanced means
for shaping trafc on the network to
decrease congestion.
iSCSI trafc also benets from the
CRC32-C instructions supported in the
Intel Xeon processor 5500 series and
the upcoming next-generation Intel
Xeon processor-based server platform
for the expandable server segments.
Data integrity checking with CRC32 has
historically been very processor-intensive,
and many IT organizations chose not to
use it. Because CRC32-C can be enabled
without undue burden on the processor,
SAN administrators can use it to improve
data integrity while accelerating
processing, relative to prior generations.
Both Windows Server 2008 and ESX 3.0
and later include support for this feature.
Architectural enhancements on these
new Intel® server platforms, including
PCI Express* Gen 2, an integrated
memory controller, and Intel® QuickPath
Interconnect, greatly increase the
scalability of 10GbE. Previous-generation
platforms did not scale beyond
approximately 17 Gbps because of
architectural limitations. Conversely, the
architecture of the Intel Xeon 5500 series
processor-based server allows it to scale
up to nearly 50 Gbps or four GbE ports.
This level of throughput scalability will
be put to good use by today’s evolving
dynamic, virtualized data centers.
Advanced Support for Virtualization
Virtualization drives the need for added
SAN connectivity to allow the advantages
that originate from remote boot, live
migration, and related virtualized usage
models. Intel Ethernet Server Adapters
offer virtualization optimizations that
directly benet iSCSI implementations.
In conventional virtualization
implementations, the hypervisor abstracts
the I/O devices and shares those
hardware resources among multiple VMs.
To route trafc coming from a shared I/O
device, a virtual switch (vSwitch) within
the hypervisor sorts incoming packets
based on the destination VM and then
delivers them accordingly. This sorting and
grouping done in the hypervisor consumes
processor cycles, negatively impacting
overall network I/O.
Virtual Machine Device Queues (VMDq)
technology, a component of Intel VT-c,
relieves the hypervisor of this overhead
and accelerates I/O trafc in virtualized
servers by performing data sorting in the
Ethernet controller. VMDq technology
makes use of multiple queues in the
network device. With VMDq, data packets
are sorted as they enter the network
adapter, and packets going to the same
destination VM are grouped together in
the same queue. The packets are then
sent to the vSwitch, which directs them
to their respective destinations. VMDq is
supported in VMware ESX and Microsoft
Hyper-V*.
Performance Results: 1 Million IOPS
Testing in Intel labs achieved one million
bi-directional average transmit-and-
receive I/O operations per second (IOPS)
at block sizes of 512 bytes,
1
without
proprietary iSCSI ofoads. Shown in
Figure 1, this result using 10GbE Intel
Ethernet Server Adapters and Intel
Xeon processor 5500 series-based
servers represents new levels of iSCSI
performance.
This level of performance, attained using
the native iSCSI initiator in Windows
Server 2008 R2, is approximately twice
that achieved using previous-generation
solutions.
2
5
Simple, Reliable Performance for iSCSI Connectivity
5