Realize Unprecedented Application Performance by Offloading
Compute-Intense Data Communication
Business Benefits
Increase Server
Efficiencies
Sustain near
line rate
bandwidth
Increase CPU
efficiency by
33%
Applications Acceleration with Mellanox
RDMA enabled Networking Solutions
Executive Summary
The expansion of business-critical
information, unstructured data, and rich
content within enterprises continues to
change storage dynamics. This trend is
driving the need for higher connectivity
speeds to keep up with applications
requirement for access to data. In order
to address these market conditions,
Microsoft and Mellanox have joined
forces to solve the storage access
problem by removing CPU overhead and
providing faster and more efficient access
to data for enterprise applications.
Microsoft SMB
The solution comes upon enhances to
Microsoft’s Server Message Block (SMB)
3.0 protocol in Windows Server 2012 R2.
The SMB protocol was extended to
include two new features, SMB Direct,
and SMB Multichannel. SMB Direct
implemented the use of various high-
speed RDMA methods to transfer large
amounts of data with little CPU
intervention. SMB Multichannel allows
file servers to use multiple network
connections simultaneously and provides
fault tolerance through automatic
discovery of network paths to dynamically
add connections as required. The addition
of these two features allows Mellanox
RDMA enabled IO adapters to deliver line-
rate performance and improve availability
by optimizing data transfer between
server and storage over standard
Ethernet. By offering built-in support for
Mellanox RDMA, Microsoft has enabled
customers to deploy storage on low-cost
file servers, while delivering higher
performance. This enables storage
that rivals costly Fibre Channel SANs in
efficiency, with lower latency and
operating cost.
Increasing Server Efficiency
Mellanox ConnectX®-4 Lx EN Ethernet
adapters support RDMA at 10/25/40/50
GbE speeds while the ConnectX®-4
network adapters offer 10/25/40/50 and
100Gb/s Ethernet connectivity and
provide the highest performance and
most flexible solution for high-
performance RDMA implementations.
RDMA provides low latency by removing
processing of the protocol stack for data
transmission from the operating system,
allowing applications to directly read and
write to remote virtual memory and to
directly exchange messages. This
drastically reduces the CPU’s involvement
in data I/O requests and reduces memory
bandwidth bottlenecks. Mellanox RDMA
technologies process all transport
protocol in the adapter hardware,
completely bypassing the host OS. This
permits the adapters to transfer data
between servers and storage with
minimum involvement from the host CPU.
As a result, SMB Direct is extremely fast
with client-to-file server performance,
almost equaling solution that use locally
attached storage. The CPU reduction in
this process leaves more cycles available
for server applications. The result is
larger numbers of hosted VMs per
physical servers, more VDI instances
running simultaneously and SQL
environments achieve high-performance
file sharing to complete queries quicker.
Data Centers efficiencies and scalability
increases in each of these instances.
SOLUTION BRIEF
Achieve over
1 Million
IOPS
Reduce
latencies by
50%
Hyper-V Use Case
The purpose of the use case is to show
the efficiency of the CPU when running
workloads within Hyper-V. The
demonstration utilized two servers, a
client and file server, that were
connected back-to-back with Mellanox
ConnectX®-4 100Gb/s Ethernet adapters
running SMB Direct over RoCE (RDMA
over Converged Ethernet). Three Micron
NVMe SSD 300GB 6G SAS adapters were
used to form an all-SSD storage pool and
IOMeter for the primary I/O generator
and benchmarking tool. To simulate a
real-world environment, different
workload were used to provide data
streams with both static and mixed I/O.
Four cores were allocated to running the
data communication task and the proof of
concept run with and without RoCE
enabled. The results demonstrated that
workloads running on Hyper-V VMs over
RoCE can achieve over 1 million IOPS of
throughput and sustain a near line rate of
100Gbps of aggregate bandwidth without
taxing the CPU. When compared against
the result with RoCE disabled, latency
decreased nearly 50% and CPU efficiency
by more than 33%. The CPU was bound
due to the requirement to process the
data transport protocol; sequence
checking, packet loss recovery, CRC
checking, and passing kernel to user space
transmission. Running the data transport
stack in the CPU steals cycles from
applications and limits overall
productivity. The Mellanox adapter frees
the CPU from this process, delivering
improved efficiencies.
RoCE Advantages
By using Microsoft’s SMB 3.0 over
Mellanox RDMA technologies,
organizations can deliver extremely high
levels of storage performance and
resiliency. RoCE requires very low CPU
utilization and takes advantage of Priority
Flow Control in Data Center Bridging
Ethernet for lossless connectivity.
Configuring Microsoft SMB with Mellanox
RDMA enabled workloads running in
Hyper-V to achieve over 1 million IOPS.
This is not a novel idea, the gaming
industry has been offloading graphic
processing for a long time. And GPU-
accelerated computing has been used to
accelerate scientific, analytics, and
engineering applications since the mid-
2000s. Adding these powerful capabilities
into your arsenal enables a
comprehensive platform that is poised to
tackle the storage demands of the
modern data center.
Figure 1. Microsoft and Mellanox 100Gb/s cloud efficiency demonstration
For more information:
Download the
Microsoft White Paper
Download a Mellanox
Case Study:
http://www.mellanox.com/related-
docs/case_studies/CS_MHA_Cloud
.pdf
http://www.mellanox.com/related-
docs/case_studies/CS_Financial_In
stitute.pdf
To learn more about
Microsoft SMB visit:
https://technet.microsoft.com/enu
s/library/hh831795(v=ws.11).aspx
More information about the
technology demonstration can
be seen in the Microsoft 2015
Ignite conference presentation
given by Claus Joergensen,
Principal Program Manager,
Microsoft
Enabling Private Cloud Storage
Using Servers with Local Disks
SOLUTION BRIEF: Applications Acceleration with Mellanox RDMA enabled Networking Solutions
350 Oakmead Parkway, Suite 100, Sunnyvale, CA 94085
Tel: 408-970-3400 • Fax: 408-970-3403
www.mellanox.com
© Copyright 2016. Mellanox Technologies. All rights reserved.
Mellanox, Mellanox logo, and ConnectX are registered trademarks of Mellanox Technologies, Ltd. Mellanox NEO is a trademark of Mellanox Technologies, Ltd.
All other trademarks are property of their respective owners.
Technology
Demonstration