2. Outline
Large Workload vMotion challenges
Enabling Multiple-NIC vMotion
Traffic Flow Considerations
QOS
NIOC
Class Based QOS on the 1000v
Real World Design Discussion
Quad 10Gb CNA / 1000v / FCoE on UCS
Questions
3. Challenges with Large Workloads
At Medtronic
78% of servers are virtualized, the low hanging fruit is already gone.
Remaining physical servers are 64GB and larger – Exchange, Oracle,
SQL, SAP Middleware
Experienced vMotion failures with large workloads on ESX 4.1
Aging vMware Hosts (3+ Years)
Requirements for new a environment
Reduced physical footprint
Support for a few guests up to 256GB (Current requests are for
128GB)
High consolidation ratio – 100+ VMs per Host
Network cable consolidation and operationalize support
4. Large Workload vMotion
Two key features of ESX 5 provide better support for vMotion of
larger workloads than previous versions
Multiple-NIC vMotion provides more bandwidth to
Motion process
More bandwidth is always better…the faster the pre-copy phase
completes the less time the guest has to dirty the pages…
Reduced time to evacuate a host going into maintenance mode
Stun During Page-Send (SDPS)
SDPS can induce small delays in processor scheduling reducing
the rate that the guest is ―dirtying‖ memory pages
Guest performance is only reduced if the guest is ―dirtying‖
memory pages faster than vMotion can pre-copy them
5. Multiple-NIC vMotion Performance
With QOS + FCoE, Without Jumbo Frames
vMotion Throughput on 10G CNA (Gb per
Second) on 1000v / UCS 6248 FI
14
12
10
8
6
4
2
0
1 10G CNA 2 10G CNA 4 10G CNA
6. Enabling Multiple-NIC vMotion
• Follow best practices and use dedicated VMKernel interfaces for
mgmt, vmotion, storage, etc…
• Create a vMotion VMkernel interface for each physical NIC you would
like to use for vMotion traffic
• For all practical purposes the vMotion VMKernel interfaces need to be
backed by the same VLAN and address within the same subnet.
• All VMKernel interfaces enabled for vMotion will be used for both
single or multiple concurrent vMotions
• Supports up to 16 interfaces with 1Gb NICS, or 4 interfaces with 10Gb
8. VMKernel to NIC Association - vDS
Create dvPortGroups before creating VMKernel adapters
Create one dvPortGroup for each physical NIC you want to
carry vMotion traffic
9. VMKernel to NIC Association – 1000v
VPC-HM with Mac-Pinning
Operates in a similar manner to vSwitch and vDS default
options, VMKernel interfaces are pinned to physical NIC
―channel-group auto mode on mac-pinning‖ is used to
enable in the ethernet (uplink) port-profile on the 1000v
―show port-channel internal info all‖ to learn pinning id
Apply the pinning-id command to pin a VMKernel interface
to a NIC
port profile type vethernet vMotionA
pinning id 1
port profile type vethernet vMotionB
pinning id 2
Verify = module vem # execute vemcmd show pinning
10. VMKernel to NIC Association – 1000v
VPC LACP
Traditional LACP based etherchannel (Active or Passive)
Upstream switch needs to support multi-chassis etherchannel
―channel-group auto mode active‖ is used to enable in the
ethernet (uplink) port-profile on the 1000v
vMotion VMKernel traffic is distributed among the member
interfaces based on the selected load balancing algorithm
―port-channel load-balance ethernet‖ to change algorithm.
If the default isn’t distributing vMotion traffic evenly try
―source-ip-vlan‖ and use consecutive IP address for the
vMotion enabled VMKernel interfaces on a host.
Use increments of 2,4,8 ports for even distribution
11. Traffic Flow During a vMotion
vCenter steps through the list of vMotion VMKernel adapters on
each host in the order they were presented to vCenter and pairs
them off
Speed mismatch will be handled by bandwidth – multiple 1Gb
NICs can be paired to a single 10Gb
There isn’t a way to reliably control which interfaces are paired
up, this could lead to vMotion traffic overwhelming switch
interconnects
A dedicated vMotion switch avoids switch interconnect issues
Multi-chassis etherchannel eliminates switch interconnect issues
If the NICs aren’t dedicated to vMotion use QOS
14. Network IO Control
Only available on the dVS (requires enterprise+ )
Has built-in resource pools for classes of system traffic
such as vMotion, Management, iSCSI, NFS
Traffic shares assign a relative importance to traffic
that is used to create minimum bandwidth reservations
on a per dvUplink basis
Only applies to outbound traffic
Limits are used to cap traffic on a per dVS basis
15. Class Based WFQ on the 1000v
CBWFQ QOS provides minimum bandwidth reservations on a per-
physical port basis
Provides built in protocol matches to classify
n1kv, vMotion, management, and storage traffic
Only applies to outbound traffic
QOS on the 1000v is a three step process
1. Define traffic classes using the class-map command
2. Create a traffic policy with the policy-map command
3. Attach the traffic policy to an interface or port-profile with the
service-policy command
FOR CBWFQ QOS EXAMPLE PLEASE DOWNLOAD THE PRESENTATION
16. Real World Design Discussion
- Four 10Gb FcOE CNA on UCS
Design Goals –
Support large workloads with up to 256Gb of RAM
Operationalize support into existing frameworks
Support both FC and NFS storage to consolidate existing farms
UCS c460 with 1Tb of RAM, two P81E dual port FCoE VICs
UCS 6248 Fabric Interconnect with 2232 FEX
40Gb VPC uplink from each fabric interconnect
Nexus 1010X / 1000v
4 vNICs and 4 vHBAs presented to ESX
Four vMotion VMKernel interfaces per host
Currently running at a consolidation ratio of 157 – 1.
Replaced 96 ESX Hosts with 8…
Succesful vMotion of an 8 way 256gb VM running SQLioSIM.
17.
18. Four 10Gb FCoE CNA on UCS
QOS Marking Policy • Apply to vethernet port-profiles
policy-map type qos class-cos1 port-profile vMotionA
description vMotion
service-policy input class-cos1
class class-default
port-profile vMotionB
set cos 1
service-policy input class-cos1
policy-map type qos class-cos2
description NFS port policy NFS
class class-default service-policy input class-cos2
set cos 2 port-policy v174
policy-map type qos class-cos4 service-policy input class-cos4
description Gold-Data port-policy ESX-Management
class class-default service-policy input class-cos6
set cos 4
policy-map type qos class-cos6
description ESX-Management
class class-default
set cos 6
19. Four 10Gb FcOE CNA on UCS
MGMT
SAP
NFS
vMotion
• vNIC QOS Policy must be set to ―host control full‖ to trust COS markings
• 1000v has no visibility to vHBAs utilization of the link
• Instead of queuing on the 1000v the UCS will Queue on the adapter and fabric
interconnect
• The ―Palo‖ adapters are reduced to three queues when placed in host control
full,
• The UCS fabric interconnect leverages the advanced QOS functions of the 5k
hardware such as virtual output queues to provide effective ingress queuing
20.
21. Key Takeaways
ESX 5 can vMotion larger guests than 4.1 with the
addition of SDPS, but more bandwidth reduces the
impact of vMotion on the guest
Consideration should be given to traffic flow when
implementing multiple-NIC vMotion, switch
interconnects can be easily overwhelmed
Dedicated vMotion adapters are best and should always
be used in 1G environments, but aren’t always practical
in 10G environments
Without dedicated adapters QOS both on the virtual and
physical switch become important
22. Questions ?
Relevant Sessions Remaining :
INF-VSP1549 - Insight Into vMotion:
Architectures, Performance, Best Practices, and Futures
INF-NET2161 - VMware Networking 2012: Enabling the
Software Defined Network
INF-NET1590 - What's New in vSphere – Networking
INF-NET2207 - VMware vSphere Distributed Switch—
Technical Deep Dive
Please complete your session surveys
heath@heathreynolds.com
24. Class Based WFQ on the 1000v
Step 1 – Classify the traffic
class-map type queuing match-any n1kv_control_packet-class
match protocol n1k_control
match protocol n1k_packet
match protocol n1k_mgmt
class-map type queuing match-all nfs-class
match protocol vmw_nfs
class-map type queuing match-all vmotion_class
match protocol vmw_vmotion
class-map type queuing match-all vmw_mgmt_class
match protocol vmw_mgmt
25. Class Based WFQ on the 1000v
Step 2 – Create a policy
policy-map type queuing uplink_queue_policy
class type queuing n1kv_control_packet_class
bandwidth percent 10
class type queuing nfs_class
bandwidth percent 25
class type queuing vmotion_class
bandwidth percent 20
class type queuing vmw_mgmt_class
bandwidth percent 10
Step 3 – Attach policy to an interface
port-profile uplink
service-policy type queuing output uplink_queue_policy