Transcript Windows Server 2012 Networking Performance and Management

Beyond
Virtualization
Windows Server 2012 offers a
dynamic, multi-tenant
infrastructure that goes
beyond virtualization to
provide maximum flexibility
for delivering and connecting
to cloud services.
The Power of Many
Servers, the
Simplicity of One
Every App,
Any Cloud
Windows Server 2012 offers
excellent economics by
integrating a highly
available and easy to
manage multi-server
platform with breakthrough
efficiency and ubiquitous
automation.
Windows Server 2012 is a
broad, scalable and elastic
server platform that gives you
the flexibility to build and
deploy applications and
websites on-premises, in the
cloud and in a hybrid
environment, using a consistent
set of tools and frameworks.
Modern Workstyle,
Enabled
Windows Server 2012empowers
IT to provide users with flexible
access to data and applications
from virtually anywhere on any
device with a rich user
experience, while simplifying
management and helping
maintain security, control and
compliance.
A two-tenant environment built with Hyper-V
in Windows Server 2012
Basic model of Resource Metering
.
.
.
.
.
.
.
.
Organize, assign, monitor and
manage static and dynamic IPv4/v6
addresses. Integration with
SCVMM
Address
space
mgmt
(ASM)
Network
discovery
.
In-box solution that
complements – and seamlessly
integrated with – MS DHCP
and DNS offerings
WS 2012
IPAM
Multiserver
mgmt
(MSM)
.
Suitable for MS DHCP/DNS
customers looking for an
economic, capable and scalable
IPAM solution
Visibility &
audit
Automatic discovery of DC,
DHCP and DNS servers, and
dynamic IP addresses in use
Centralized configuration and
update of MS DHCP/DNS servers
Track and audit changes and
provide real-time view of
status
Win 8 (RSAT)
& WS 2012
Agentless architecture
WS 2012 in-box
WS08; WS08 R2 & SPs; WS 2012
Distributed deployment, scale, and DR
• Processor. One physical processor, which can consist of
•
•
•
•
•
one or more nodes. A physical processor is the same as
a package, a socket, or a CPU.
Non-uniform memory architecture (NUMA) node. A
set of logical processors and cache that are close to one
another.
Core. One processing unit, which can consist of one or
more logical processors.
Logical processor (LP). One logical computing engine
from the perspective of the operating system,
application, or driver. In effect, a logical processor is a
thread.
Kernel Group. A set of up to 64 logical processors.
Affinity. A preference indicated by a thread, process, or
interrupt for operation on a particular processor, node,
or group.
• Latency. The time required for an operation to complete. Lower is
better.
• Scalability. The ability to adapt to increasing demand on system
resources. Higher is better.
• Throughput. The amount of data transferred or processed in a
given time period. Higher is better.
• Path Length. The number of CPU cycles divided by the throughput.
Lower is better.
• Variability / Jitter. Fluctuation in throughput and/or latency. Lower
is better.
Receive Segment Coalescing (RSC)
• RSC can help by coalescing multiple
inbound packets into a larger buffer and
effectively reduce per packet CPU cost
• Preliminary data shows up to a 20%
reduction in CPU cycles depending on
the workload
• All inbox 10G network adapters support
the feature and have it on by default
RSC: Multiple Inbound Packets
Coalesced to reduce
CPU load
Receive Side Scaling (RSS)
• Use RSS to distribute the receive
•
network traffic either across multiple
processors or multiple NUMA nodes
IT Pros should use new Power Shell
cmdlets to control RSS behavior based
on their workload characteristics
Static
Dynamic
NUMA
aware
Profile 3
Profile 2
Non
NUMA
Profile 1
Profile 0
(like WS08 R2)
(default)
Receive Side Scaling (Continued)
Node 1
Node 2
Node 3
0
1
2
3
0
1
2
3
queues
Node 0
Incoming
Packets
NIC with 8 RSS queues
New RSS profiles improve scalability on multi-node servers by
distributing TCP/UDP receive traffic across nodes
Single Process
Multiple Processes (2 and 4)
50%
2 Nodes (20LP) 4 Nodes (40LP)
115%
WS08R2
Win2012
Throughput
Throughput
201%
Hardware: 40LPs, 4 NUMA
nodes , 10GigE, 128GB RAM
Workload: mixed Web
scenario
43%
2 Nodes (20LP) 4 Nodes (40LP)
WS08R2
Win2012
Note: Measurements reflect
release defaults
Dynamic Virtual Machine Queues (VMQ)
• Use VMQ to dynamically distribute the receive network traffic
•
across multiple VMs based on destination and network load
Most inbox 10Gig Ethernet adapters support the feature which is
enabled by default
Network IO path without VMQ
Network IO path with VMQ
Registered I/O (RIO)
• Registered I/O (RIO) delivers low
latency by “pinning” the application
memory and reduces CPU cost
• RIO improves predictability with
message latency (jitter)
• Example APIs: RIORegisterBuffer
and RIOReceive
TCP Loopback Fast Path
• Customer should use new IOCTL
SIO_LOOPBACK_FAST_PATH socket
option to take new shorter path
• Both sides of connection must set
new option but can’t have
SO_OOBINLINE,
TCP_EXPEDITED_1122, or
TCP_STDURG socket options
Winsock Application
(E.g. Using localhost as TCP destination)
user
kernel
Win08 R2
Loopback
Path
AFD
Windows Server
2012 Loopback Path
TCP/IP
NDIS
Windows 2008 R2
New: Shorter Path
• IT Pro wants to fine tune their system for low
latency application and minimize jitter
• Best Practices include:
• Disable C-states from BIOS
• Set Power Policy to High Performance
•
•
•
•
•
•
Use powercfg utility
Disable NIC Interrupt Moderation setting
Disable Flow Control
Disable Receive Side Scaling
Disable Receive Segment Coalescing (RSC)
Use thread affinity in Task Manager to partition
work by nodes
• Check out http://technet.microsoft.com/enus/library/hh831415.aspx
Server Performance Advisor (SPA) 3.0
Original Setting is derived from
system tuning, empirical evidence
New Performance Monitor Counters
•
RDMA Activity
PowerShell commands
•
IPsecOffload
Lso
Rss
NetAdapter.Get-NetAdapterRsc
Vmq
SrioVf
PowerShell commands (Continued)
NetAdapterEncapsulatedPacketTaskOffload
NetAdapterIPsecOffload
Enable-NetAdapterLso
Enable-NetAdapterRdma
Enable-NetAdapterRsc
When do you use each feature?
Performance Loopback Registered
Metric
Fast Path I/O (RIO)
Lower Endto-End
Latency
X
Higher
Scalability
Large
Send
Offload
(LSO)
Receive
Segmentation
Offload
(RSC)
Virtual
Machine
Queues
(VMQ)
X
X
Higher
Throughput
X
X
X
X
Lower Path
Length
X
X
X
X
Lower
Variability
Receive
Side
Scaling
(RSS)
X
X
X
X
X
Remote Single
DMA
Root I/O
(RDMA) Virtual
(SR-IOV)
X
X
X
X
X
X
Subtitle color
www.microsoft.com
Windows PowerShell Crash Course
WSV307: Windows Server 2012 IP Address Management
WSV314: Windows Server 2012 NIC Teaming and Multichannel Solutions
VIR303: An Overview of Hyper-V Networking in Windows Server 2012
VIR305: Hyper-V Network Virtualization for Scalable Multi-Tenancy in
Windows
VIR307: Get Hands-on with the New Hyper-V Extensible Switch in Windows
Server 2012
WSV14 – HOL: Managing Your Network Infrastructure with IP Address
Management
WSV24 – HOL: Managing Network Infrastructure with Windows Server 2012
#TEWSV304
Hands-On Labs
DOWNLOAD
Windows Server
2012 Release
Candidate
DOWNLOAD
Microsoft System
Center 2012
Evaluation
microsoft.com/windowsserver
microsoft.com/systemcenter
http://europe.msteched.com
www.microsoft.com/learning
http://microsoft.com/technet
http://microsoft.com/msdn
http://europe.msteched.com/sessions