hp discover 2012 tb#3258: the benefits and right practices of 10gbe networking with vmware vsphere 5...
TRANSCRIPT
HP Discover 2012
TB#3258: The benefits and right practices of 10GbE networking with VMware vSphere 5
June 6, 20122:45PM
© 2012 Emulex Corporation2
© 2012 Emulex Corporation2
Agenda
The Transition…
What’s new with VMware vSphere 5
The design considerations
Deployment scenarios
Summary
The Transition
© 2012 Emulex Corporation4
© 2012 Emulex Corporation4
A website focused on technical content
A consortium of technical whitepapers of actual work done in Emulex Labs or Partner Labs
A resource for how-to content, blogs, application notes, technical videos and technical webcast recordings
Recently selected as one of the Best Technical websites by Network Products Guide – May 2012
© 2012 Emulex Corporation5
© 2012 Emulex Corporation5
100Mb 1Gb 10GbUTP Cat 5 UTP Cat 5
SFP Fiber
10MbUTP Cat 3
Mid 1980’s Mid 1990’s Early 2000’s
CableTransceiver
Latency (link)Power
(each side)DistanceTechnology
Twinax ~0.1ms~0.1W< 10mSFP+ Direct Attach
MM 62.5mmMM 50mm
~01W82m
300mSFP+ SR
short reach
Ethernet Physical Media
SFP+ Cu, FiberX2, Cat 6/7
Late 2000’s
© 2012 Emulex Corporation66
*FDDI-grade MMF
10GbE PHY Standards
10GBASE- Media Reach Module Cost
LR SMF 10km High
SR MMF 26m* Medium
LX4 MMF/SMF 300m* High
CX4 Twinax 15m Low
LRM MMF 220-300m* Medium
T UTP 30-100m Not applicable
L – 1310nm LaserS – 850nm LaserX – Uses 8B/10B encodingR – Uses 64B/66B encodingT – Unshielded Twisted PairC – Copper (Twinax), transmits over 4-lanes in each direction4 (as in LX4) – represents information is transmitted over 4 wavelengths using a coarse Wavelength Division Multiplexing
© 2012 Emulex Corporation7
© 2012 Emulex Corporation7
Upgrade Current Network or Greenfield Build
If upgrading current network– Consider bandwidth requirement of servers– Implement 10GbE uplink for 1Gb devices– High priority devices attach directly to 10GbE
Greenfield Data Center 10Gb– 10GbE to Top of Rack– Edge / Core Switching at 10Gb– 10GbE Convergence NIC/iSCSI/FCoE Throughout
© 2012 Emulex Corporation8
10GbE Enterprise Class iSCSI/LAN – Optimized Server Virtualization
1GbE LOM and Multiple 1Gb Adapters10GbE LOM and 10Gb Dual-Port
Adapter
LAN 3
LAN 4
vMotion
Administration
iSCSI 2
iSCSI 1
LAN 1
LAN 2
iSCSI / LAN
vMotion
Administration
iSCSI / LAN
What’s New in vSphere Networking with 10GbE
© 2012 Emulex Corporation10
© 2012 Emulex Corporation10
VMware Networking Improvements
vMotion• Throughput improved significantly for single vMotion
– ESX 3.5 – ~1.0Gbps– ESX 4.0 – ~2.6Gbps– ESX 4.1 – max 8 Gbps– ESX 5.0 – Multiple 8+ Gbps connections
• Elapsed reduced by 50%+ on 10GigE tests.
Tx Worldlet• VM – VM throughput improved by 2X, to up to 19 Gbps• VM – Native throughput improved by 10%
LRO (Large Receive Offload)• Receive tests indicate 5-30% improvement in throughput• 40 - 60% decrease in CPU cost
© 2012 Emulex Corporation11
Configure jumbo MTU VM virtual adapter – change jumbo packet size value to 9k ESX host – change vswitch MTU to 9k NICs at both ends and all the intermediate hops/routers/switches must also support jumbo frames
Example: Configure 9k packet size on W2k8 VM virtual adapter
Jumbo MTU with 10GbE
© 2012 Emulex Corporation12
Configure jumbo MTU– On ESX host, change vSwitch1 MTU to support jumbo frames up to 9000
Example: Configure 9k MTU to vSwitch1
Jumbo MTU with 10GbE
© 2012 Emulex Corporation13
© 2012 Emulex Corporation13
Jumbo Frames even more important with 10GbE
Performance Improvement in iSCSI and NFS read/write throughput using Jumbo Frames on 10GbE
Protocol Read Throughput Write Throughput
SWiSCSI +11% +39%
NFS +9% +32%
64KB Blocksize
© 2012 Emulex Corporation14
NetQueues
What are NetQueues? Provides 8 queues on each 10G port for incoming network traffic Offloads incoming packet sorting/routing functions from hypervisor to the
adapter Frees up CPU resources Improves platform efficiency and overall system performance Traffic is balanced across multiple CPU cores for processing using MSI-x Message Signaled Interrupts (MSI-x) support
o Doubles the # of device interrupts generated o Balances interrupts across all CPUs in a Symmetric Multiprocessing
(SMP) platform Optimized throughput between multiple virtual machines and a physical
network Best benefit with jumbo MTU on vSwitch, VNIC, physical switch
© 2012 Emulex Corporation15
NetQueues
Without NetQueues Enabled – Only a single TX and RX queue for processing all network traffic generated by multiple virtual machines.
Enable NetQueues in ESX Host or use vSphere Client
Configuring NetQueues on ESX Host:
© 2012 Emulex Corporation16
Performance Benefits Enabling NetQueues and Jumbo MTU
1500 MTU, 8VMs no NetQueue Receive Thru = 7.36 Gbps 9000 MTU, 8VMs NetQueue enabled Receive Thru = 9. Gbps
• Throughput is increased by 2.50 Gbps – ~ NIC Line Rate
• Reduces average CPU utilization/used
• 35 - 50% reduction in Virtual Machine CPU utilization
Throughput Across VMs
© 2012 Emulex Corporation17
© 2012 Emulex Corporation17
Network Traffic Management with 10GbE
Dedicated NICs for various traffic types e.g. VMotion, IP Storage
Bandwidth assured by dedicated physical NICs
FT VMotion NFS
vSwitch
TCP/IP
iSCSI FT VMotion NFS
vSwitch
TCP/IP
iSCSI
1GigE 10 GigE
Traffic typically converged over two 10GbE NICs
Some traffic types & flows could dominate others through oversubscription
Traffic Types compete.
Who gets what share of the vmnic?
© 2012 Emulex Corporation18
NetIOC: Guaranteed Network Resources
NetIOC disabled
NetIOC enabled
NFS, VM, and FT traffic take a dip during vMotion
Shares & Limits as Configured
NFS, VM, and FT traffic not affected by concurrent vMotion
© 2012 Emulex Corporation19
© 2012 Emulex Corporation19
802.1p Tagging (QoS)
802.1p is an IEEE standard for enabling Quality of Service at MAC level.
802.1p tagging helps provide end-to-end Quality of Service when:• All network switches in the environment treat traffic according to the tags• Tags are added based on the priority of the application/workload
You will now be able to tag any traffic flowing out of the vSphere infrastructure.
© 2012 Emulex Corporation20
© 2012 Emulex Corporation20
VMware vMotion
Benefits of vMotion– Eliminates planned downtime
– Enables dynamic load balancing
– Reduces power consumption
– Essential to managing the virtualized datacenter
Challenges– Larger VMs require more time to
vMotion
– Larger environments require more vMotions to remain balanced
© 2012 Emulex Corporation21
vSphere 5 vMotion – Multi-NIC support
Tim
e in
se
con
ds
1 VM 2 VMs0
10
20
30
40
50
60
70
vSphere 4.1vSphere 5.0vSphere 5.0 (2 NICs)25 to 30% Better
50% BettervSphere 5.0, out of box configuration provides 25% to 30% reduction in vMotion Time
Using 2 NICs can provide linear scaling and a 50% reduction in vMotion time
Elapsed reduced by 50%+ on 10GbE tests
© 2012 Emulex Corporation22
© 2012 Emulex Corporation22
VMware Fault Tolerance
Single identical VMs running in lockstep on separate hosts
Zero downtime, zero data loss failover from hardware failures
No complex clustering or specialized hardware
Single common mechanism for all applications and OS-es
10GbE Benefits
• 20% Better VM Performance
• Easier to Manage Less NICs required
• 10GbE is Highly Recommended for FT
© 2012 Emulex Corporation23
vSphere 5 – Networking Improvements (SplitRXmode)
Greatly reduce packet loss for Multicast processing
vSphere 5 introduces a new way of doing network packet receive processing • Splitting the cost of receive packet processing to multiple contexts.• Each VM we can specify whether we want to have receive packet processing in
the network queue context or a separate context.
Results when we had more than 24 VMs powered on Without splitRxMode up to 20% packet loss. Enabling splitRxMode less than 0.01% packet loss.
Enable SplitRxMode on a per VNIC basis Editing vmx file of VNIC with ethernetX.emuRxMode = "1" for the ethernet device.
Note: This is only available with vmxmet3.
© 2012 Emulex Corporation24
© 2012 Emulex Corporation24
Network Latency
VM round trip latency overhead 15-20 microseconds
DirectPath reduces round trip latency by 10 µs
How to reduce Network Latency– Consider disabling power management
– Consider disabling physical & virtual NIC interrupt coalescing
– Reduce Contention on the physical NIC
– Utilize 10Gb NICS
– Utilize VMXNET3 Para virtualized device driver
– Utilize Directpath (if really needed)
http://www.vmware.com/files/pdf/techpaper/network-io-latency-perf-vsphere5.pdf
© 2012 Emulex Corporation25
Networking & 10GbE Best Practices
Turn on Hyper-threading in BIOs
Confirm that the BIOS is set to enable all populated sockets for all cores
Enable “Turbo Mode” for processors that support it
Confirm that hardware-assisted virtualization features are enabled in the BIOS
Disable any other power-saving mode in the BIOS
Disable any unneeded devices from the BIOS, such as serial and USB ports
For Windows VMs use VMXNET3 Enhanced Virtual Adapters for best performance
Adjust Network Heap Size for excessive network traffico By default ESX server network stack allocates 64MB of buffers to handle network datao Increase buffer allocation from 64MB to 128MB memory to handle more network datao To change Heap Size ESX Host: # excfg-advcfg –k 128 netPktHeapMaxSize
© 2012 Emulex Corporation26
© 2012 Emulex Corporation26
Networking & 10GbE Best Practices Cont’d
Be Mindful of Converged networks, storage load can effect network– Use NIOC to balance and control network workloads
Use Distributed switches for cross ESX host Network Convenience, no significant performance impact for DvSwitch vs. vSwitch
– DvSwitch Needed for NIOC and advance network shaping
Utilize the latest NIC Features– Jumbo Frames for iSCSI & NFS workloads
– Utilize 10Gb Ethernet hardware
Utilize NIC Teams for fail over (HA) and NIC load balancing– Use Multiple NICs for improved vMotion Speeds
Keep an eye out for Packet Loss and Network Latency
Design Considerations
© 2012 Emulex Corporation28
© 2012 Emulex Corporation28
Design Considerations
Separate Infrastructure Traffic from VM Traffic– VMs should not see infrastructure traffic (security violation)
Method of Traffic Separation– 1. Separate logical networks (VLANs)
• Create one VDS, connect all pNICs• Create portgroups with different VLANs, place vmknic and VM’s vNics on
different portgroups• 2 pNics are sufficient
– 2. Separate physical networks • Create one VDS for each physical network• Create portgroups, put vmknic and VM’s vNics on portgroups from different
vswitches• Requires at least 4 pNICs (2 per network for redundancy)
© 2012 Emulex Corporation29
© 2012 Emulex Corporation29
Design Considerations – Cont’d
Avoid Single point of Failure– Connect two or more physical NICs to an Uplink Port group– Preferably connect those physical NICs to separate physical switches
Understand your Virtual Infrastructure traffic flows– Make use of NetFlow feature to monitor the traffic flows over time– Use the data to come up with appropriate traffic shares to help NIOS
configuration
Prioritize traffic for important workloads and infrastructure traffic
– Use QoS tagging
© 2012 Emulex Corporation30
© 2012 Emulex Corporation30
Design Considerations – Virtual Infrastructure Traffic Types
© 2012 Emulex Corporation31
© 2012 Emulex Corporation31
Design Considerations – Traffic Characteristics
/ NFS
Deployment Scenarios
© 2012 Emulex Corporation33
© 2012 Emulex Corporation33
Deploying 10GbE
Customer Infrastructure– Needs– Blade Server Deployment
• Two – 10Gb Interfaces– Rack Server Deployment
• Two – 10Gb Interfaces
Physical Port Configurations– Multiple 1Gb NICs– Two 10Gb NICs
Physical Switch Capabilities– Switch clustering– Link State tracking
© 2012 Emulex Corporation34
© 2012 Emulex Corporation34
Rack Server Deployment – 10GbE Interface
© 2012 Emulex Corporation35
© 2012 Emulex Corporation35
Static- Port Group to NIC mapping
© 2012 Emulex Corporation36
© 2012 Emulex Corporation36
Dynamic – Use NIOC, Shares and Limits
Need Bandwidth information for different traffic types– Consider using NetFlow
Bandwidth Assumption– Management – less than 1Gb– vMotion – 2Gb– iSCSI – 2Gb– FT – 1Gb– VM’s – 2Gb
Share Calculation– Equal shares to vMotion, iSCSI and Virtual Machine– Lower shares for Management and FT
© 2012 Emulex Corporation37
© 2012 Emulex Corporation37
Dynamic – Use NIOC, Shares and Limits
© 2012 Emulex Corporation38
© 2012 Emulex Corporation38
Blade Server Deployment – 10GbE Interface
© 2012 Emulex Corporation39
© 2012 Emulex Corporation39
Static – Port group to NIC mapping
© 2012 Emulex Corporation40
© 2012 Emulex Corporation40
Dynamic – Use NIOC, Shares and Limits
© 2012 Emulex Corporation41
© 2012 Emulex Corporation41
Pros and Cons
Pros• Deterministic traffic allocation• Administrators have control over which
traffic flows through which uplinks• Physical separation of traffic through
separate physical interfaces
Cons• Underutilized I/O resources• Higher CAPEX and OPEX• Resiliency through Active-Standby Paths
Static
© 2012 Emulex Corporation42
© 2012 Emulex Corporation42
Pros and Cons
Pros• Better utilized I/O resource through traffic
management• Logical separation of traffic through VLAN• Traffic SLA maintained through NIOC shares• Resiliency through Active-Active Paths
Cons• Dynamic traffic movement across physical
infrastructure need all paths to be available to handle any traffic characteristics.
• VLAN expertise
Dynamic
© 2012 Emulex Corporation43
Emulex is HP’s Leading I/O Supplier
Partnership Across HP’s ESSN Business Units
Broadest Offering Across Operating Systems and Form Factors
Business Critical Systems
NetworkingIndustryStandardSystems
StorageDivision
Enterprise Server, Storage, and Networking
Superdome210GbE uCNA10GbE NIC4/8Gb FC
8Gb/10GbE Combo
Integrity BladeTiger LOM
10GbE uCNA10GbE NIC
8Gb FCrx2800
10GbE uCNA10GbE NIC
ProLiant DL10GbE uCNA
ProLiant ML10GbE uCNA
ProLiant SL10GbE uCNA
ProLiantBladesystemsLOM-Down (G7)
BLOM (Gen8)10GbE uCNA
10GbE NIC, 8Gb FC
5820
P10000
x9000D2D
P4000
10GbE uCNA16Gb FC
© 2012 Emulex Corporation44
© 2012 Emulex Corporation44
Resources
Emulex Implementers Lab: www.implementerslab.com– Website containing technical whitepapers, application notes, videos and
technical blogs.
HP VMware Solutions www.hp.com/go/vmware
VMware Technical Resource Center: http://www.vmware.com/resources/techresources/
www.hp.com/go/ProLiantGen8
© 2012 Emulex Corporation45
© 2012 Emulex Corporation45
Summary
Migrations from 1GbE to 10GbE deployments do require new hardware but mostly all servers and top of rack switches being deployed are now 10GbE
VMware Distributed Virtual Switch enables data center like infrastructure
VMware ESXi 5.0 complemented by 10GbE adapters provided by Emulex enable the tools and technologies required to successfully deploy 10GbE networks
HP adapters provided by Emulex deliver all the necessary technology to support VMware vSphere 5.0 10GbE network deployments
© 2012 Emulex Corporation46
© 2012 Emulex Corporation46
Emulex Breakout Sessions
Wednesday, June 6th – 2:45pm– The benefits and best practices of using 10Gb Ethernet with VMware vSphere 5.0 (session
#TB3258) Join Emulex and VMware for a technical discussion to help understand the effect of transitioning from 1Gb to 10Gb Ethernet. It’s no longer about speed and feeds as it requires knowledge of new network architectures and management tools in order to setup and monitor bandwidth allocation and traffic prioritization. Storage and Network architects need to have an understanding of new storage technologies, and networking features VMware brings with VMware ESXi 5.0.
– Speaker: Alex Amaya (Emulex), and VMware
Thursday, June 7th – 11:15am– Converged I/O for HP ProLiant Gen8 (session #BB3259 – Thursday, June 7th 11:15am) Join this interactive
session with IT Brand Pulse to understand the benefits of HP converged networking, by exploring a customer’s real-world deployment. HP has just unveiled its next generation ProLiant Gen8 platform that accelerates server market transformation – in addition to being self-sufficient, this new platform is enabled with industry-leading I/O making them incredibly fast and flexible. The HP ProLiant Gen8 platform delivers a converged I/O that includes FlexFabric, Flex-10, iSCSI and Fibre Channel over Ethernet.....at the price of 10GbE.
– Speaker, Frank Berry, CEO of IT Brand Pulse
Get a Starbucks Card! Enter to win a $100 Visa Gift Card!
Enter to win the 2012 Ducati Monster Motorcycle!
© 2012 Emulex Corporation47
© 2012 Emulex Corporation47
Emulex Booth – Demos, & Grand Prize
Management
OneCommand® Manager
Emulex OCM vCenter plug-in
8GFC HBAs – stand up and mezzanine
HP Converged Infrastructure – Path to the Cloud
• FCoE• iSCSI• 10GbE Ethernet
GRAND PRIZE
2012 Ducati Monster
Requirements for bike entry:
Listen to a booth demo
Participate in an Emulex breakout session
Announce winner on 6/7 at 1:30 PM!!!!!
© 2012 Emulex Corporation48