November 2008Leonid.Grossman@neterion.com

Direct PCI Function assignment eliminates the performance overhead of virtualization – while preserving virtualization benefits like migration, and leaving control over privileged operations with Dom0.PCI Function assignment adoption is not gated by SR IOV support – depending upon hardware design, it can work with both IOV-aware and with legacy kernels.With PCI functions implemented as complete netdevinterfaces, no split netfront/netback driver model is required – native Linux or Windows drivers can run in guests. This allows leveraging work done for the Native OSs◦ Features like Multi-Queue support, Bonding, etc.◦ Driver certification and distribution.

Performance, Performance, Performance.◦ CPU% overhead of virtualizing 10GbE can be 4x-6x,

depending upon networking workloads. ◦ Running benchmarks at 10GbE line rate in a guest is

possible with multi-queue solutions – but at the expense of driver complexity and additional cpu% overhead.

Use of single “native” drivers in a guest, regardless of which Hypervisor is deployed (or none).◦ The approach offers number of certification and

distribution benefits, as well as rich feature set of a native stack. Driver “sees” the same resources.

Incumbent solutions (emulation and para-virtualization) – are further profiled in backup slides.

SR IOV and MR IOV specifications are released by PCI SIG this year.A new generation of SR-IOV compliant NICs (like Neterion 10GbE x3100) are being released in early 2009 with standard support for multiple PCI Functions:◦ From the system prospective, each function is an independent

NIC.◦ The functions typically share an external physical port and PCI bus

– the rest of NIC resources can be dedicated.Enabled by standards based IOMMUs◦ PCI-SIG standardizes the IOMMU with SR-IOV.◦ Pre-IOMMU deployment possible for prototyping

Multi-Port GbE NICs.◦ Issues range from physical port/cables limitations to

bandwidth limitations.10GbE NICs with SR IOV VFs implemented as a subset of a NIC not as a complete NIC – for example, with VFs implemented as Level 1 or 2 queue pair.◦ Queue pair is a good way to scale across multiple cpu cores within

single OS image, but not across multiple OS imagesMore detailed analysis is included in backup slides

This presentation is not a call to add support for multi-function NICs in Xen.◦ Because the support is already there.Xen, and the various GOSs, already have almost everything they need to support multi-function NICs.◦ Xen has PCI Function Delegation.◦ Xen has migration.◦ GOSs support bonding/teaming drivers.◦ GOSs support PCI device insertion/removal.◦ An example of configuring multi-function x3100 10GbE NIC for

direct access and migration is included in backup slides.

NIC presents itself as multiple PCI functions.◦ Xen can assign as many to each guest as it wants to.Still needs to deal with more VMs than direct PCI Functions◦ Correct, only I/O intensive VMs need a dedicated function.“Needing” a h/w specific driver in guest◦ That’s not a bug. That’s a feature; a native feature-rich, certified

driver is already included in a guest.Full L2+ switch functionality on NIC◦ No need for a “full switch”, only raw frame forwarding services are

required. The functionality is similar to a soft switch in Dom0.Relies on an Address Translation Solution◦ IOMMU is just the most likely solution.

Act as a configured switch under Don0 privileged control.◦ Forwarding may be statically controlled, not learned.Direct incoming frames to the correct PCI Function.Provide internal VNIC-to-VNIC frame forwarding.Provide VNIC to external port forwarding:◦ Can prevent forged source addresses.◦ Can enforce VLAN membership.Device domain participation is optional, but Dom0 can control things it already knows (Vlan participation, etc)More details included in Backup slides.

Each Guest Driver directly controls a Virtual NIC.Each Native Driver manages only it’s VNIC.Privileged (Device Domain) Driver also manages the shared resources.◦ Or whatever portion of

them it wants to manage.

Some of the resources supported by current Neterion x3100 10GbE drivers◦ VF MAC addresses◦ VF VLAN membership◦ VF promiscuous mode◦ Bandwidth allocation between VFs◦ Device-wide errors◦ Link statistics◦ Link speed and aggregation modes

Migration Support using Guest OS Services

Frontend/Backend supplies excellent migration already◦ But requires a Hypervisor

specific frontend driver.Because it is the only universally supported solution it plays a critical role in enabling migration.

Frontend/Backend is kept in place and is always available.Direct Assignment is used for most important Guests◦ Each multi-function device will have a limit on how

many guests it can directly support.Native Driver talks directly to NIC through its own PCI function, if enabled.Bonding Driver uses frontend/backend if direct NIC is not available.More details on migration are included in Backup slides.

ServerBlades

Internal EthernetSwitch

360Gb/sec adapter aggregate bandwidth

40-80Gb/sec‘egress’bandwidth

Server Blade Chassis

External EthernetFabric

ServerManagement

Internal Ethernet SwitchManagement

Ethernet ‘Fabric’Management

Unclear management responsibility:Ethernet switches inside the Server domain

Mezz

MezzMezz

MezzMezz

MezzMezz

MezzMezz

MezzMezz

MezzMezz

MezzMezz

Mezz

Fabric Module

No link-layerflow control

Serious CongestionIssues

InteroperabilityConcerns

Non hot-swappableI/O

Extra heaton each blade

ServerBlades

Internal PCIeSwitch

360Gb/sec adapter aggregate bandwidth

20-40Gb/sec‘egress’bandwidth

Server Blade Chassis

External EthernetFabric

ServerManagement Ethernet Fabric’

Management

PCIePCIe

PCIePCIe

PCIePCIe

PCIePCIe

PCIePCIe

PCIePCIe

PCIePCIe

PCIePCIe

E3100E3100

Clear management responsibility:Servers and Network are managed independentl

Fabric Module

No InteroperabilityConcernsPCIe link-layer

flow control

E3100 10GbE ASIC and PCIe switch manage congestion

Hot-swappableI/O Module

RemoveHeat/powerFrom blades

Direct PCI Function assignment eliminates the performance overhead of virtualization – while preserving virtualization benefits like migration, and leaving control over privileged operations with Dom0.PCI Function assignment adoption is not gated by SR IOV support – depending upon hardware design, it can work with both IOV-aware and with legacy kernels.With PCI functions implemented as complete netdevinterfaces, no split netfront/netback driver model is required – native Linux or Windows drivers can run in guests. This allows leveraging work done for the Native OSs◦ Features like Multi-Queue support, Bonding, etc.◦ Driver certification and distribution.

Multi-port NIC has multiple physical ports.◦ Or possibly simply multiple NICs.

Each port is a distinct PCI function.◦ Each PCI function can be directly assigned to a Guest.◦ Provides benefits of Native Driver usage

No virtualization penalty.Full features of native mode available.Single driver in the OS image.

But an entire port is a big thing to assign◦ Not enough space for physical ports and cables.◦ No granularity on assignment.◦ Bandwidth wasted if the assignee has nothing to send.

The same mechanisms that enable direct assignment of stand-alone PCI devices also enable direct assignment of PCI functions that share a port.◦ MSI-X Mapping.◦ Page Mapping◦ Config Space Mapping.◦ Ability to use Native Drivers.Only the issues related to the shared resources need to be solved◦ Most do not require any additional management from Xen.

A Multi-queue NIC provides multiple independent queues within a PCI function.Native OS can use multiple queues itself◦ CPU Affinity, QoS, Ethernet Priorities.DomD can utilize Guest Specific queues:◦ But his is not true device assignment.

Backend must validate/translate each request (WQE/Transmit Descriptor).Does not enable vendor’s native driver.

Which already knows how to use multiple queues.

Limited functionality queues can be set up to “Queue Pairs” in Guest (or even Guest applications)Creation/Deletion/Control still managed through the Device Domain.Hardware specific code required◦ In both the Guest and Device Domain.

Single-threaded control means entangled control.◦ Need to separate control-plane work done for a guest before it

can be migrated.

Multi-function NICs present each external port as though it were a multi-queue NIC◦ But for each of multiple PCI functions.◦ Each PCI function is virtual – but that term is reserved.

Each PCI function can be directly assigned◦ Allowing Native Drivers to be used as is.◦ Allowing support of large number of Guests with a compact

number of ports.◦ Enabling existing support of multi-queue capabilities already in

the Guest OS.

RDMA-style NIC

DomU Device Domain

Queue Pair

Packet Input Routing / Output Mixing

External Port External Port

VNIC

“Verbs”Native Driver

Frontend Backend

Bridging

Application Backend

DomU

“Verbs

Frontend Application

Queu Pair

Page Protection is not enough◦ Full Function Assignment requires true isolation of each PCI

Function.Protecting Pages is not enough.If a bad configuration register can hang the device then the functions are not truly independent.

◦ Some devices can only support direct Fastpaths.A direct fastpath does not address Driver Distribution Issues – It still requires two drivers◦ One when virtualized◦ One when running in native mode.Single-path for Slowpath Control means entangled Slowpaths◦ Untangling for migration not guaranteed to be easy.RDMA users complain about cost of slowpath operations. ◦ Virtualization will only make it worse.

Availability:◦ Is the driver installed in the Guest OS image?Efficiency:◦ Does the driver interfaces efficiently with the NIC?

Migration:◦ Can Guests using this Driver be migrated?

Flexibility:◦ Can new services be supported?

Availability:◦ Excellent, NICs to be emulated are selected based on widespread

deployment.

Performance:◦ Terrible.

Migration:◦ not a problem.

Flexibility:◦ None. You’re emulating a 20th century NIC.

Availability:◦ Good. But there is a lag problem on which frontend has made it

into the OS distribution.

Performance:◦ Tolerable.Migration:◦ not a problem.Flexibility:◦ New features require extensive collaboration.

Availability:◦ Excellent. The same driver is used whether running natively or

under any Hypervisor.◦ NIC vendors already deal with OS distributions.

Performance:◦ Same as native.Migration:◦ Not really a problem, details to follow.

Flexibility:◦ Same as native.

Multi-queue is a valuable feature◦ But it does not really compensate for being a Single PCI Function

Device.

Multi-function NICs are multi-queue NICs◦ But each queue is owned by a specific PCI Function.◦ It operates within the function specific IO MAP

Allowing GOS to communicate GPAs directly to the NIC.

Each PCI Funciton has its own◦ MSI-X.◦ PCI Config space.◦ Function Level Reset.◦ Statistics.

No performance penalty◦ GOS Driver is interacting with Device the same way it would

without virtualization.There is Zero penalty to the host.Multi-function NICs offload the cost of sharing.

◦ Frontend/Backend solutions always cost more:Address translation has non-zero cost. Copying even more.Latency penalty unavoidable.An extra step cannot take zero time.

Can support ANY service supported by the Native OS.◦ because the Native OS Driver sees the same resources.

This is not a bug. It is a feature.There already is a device specific driver in the Guest OS image.The vendor worked very hard to get it there.◦ And to get it tested, and certified.◦ There is already a distribution chain

Which customers strongly prefer.◦ It already integrates hardware capabilities with Native OS

requirements.

“Soft Switches” have traditionally performed switch-like services within the Device Domain.◦ But generally have not been fully compliant 802.1Q Bridges.◦ For example most do not enable Spanning Tree.◦ And in fact that is generally a good thing.

They don’t have any real information to contribute to the topology discovery. They’d just slow things down.

The Dom0 Soft Switch will still be part of the solution◦ It must support VMs not given direct assignments.◦ Some co-ordination between the two “Switches” is desired.

At minimum these services acts as a configured switch under privileged control.◦ Forwarding may be statically controlled, not learned.

Directs incoming frames to the correct PCI Function.Provides internal VNIC-to-VNIC frame forwarding.Provides VNIC to external port forwarding:◦ Some form of traffic shaping probably required.◦ Must prevent forged source addresses.◦ Must enforce VLAN membership.

Virtual Links cannot be monitored◦ Need to know where frames really came from.

MF-NIC MUST prevent spoofing.Validating Local MAC Address is enough◦ IP and DNS can be checked using existing tools given a validated

MAC address.

Only privileged entities can grant permission to a PCI Function to use a given MAC Address.

Manufacturer◦ Each directly assigned Ethernet function is a valid

Ethernet device.It has a manufacture assigned globally unique MAC Address.No dynamic action is required to enable use of this address.

Zero configuration conventions for VLANs:◦ Membership in default VLAN is usually enabled even without

network management (usually VLAN 1).

Both network management and Xen already assign MAC Address and VLAN Memberships to Ethernet Ports.Xen can perform this function itself, or enable network management daemons in a designated domain to control the multi-function NIC’s “switch ports”.In either cases, the “port configuration” is a privileged operation through the PCI Function driver.

Neither the Hypervisor or Dom0 needs to be involved.◦ The Shared Device and its PCI Function Drivers can implement

their own solution.◦ The device already knows how to talk with each VF Driver, and

what VFs are active.

Dom0 can control things it already knows:◦ MAC Address of VNIC ports.◦ VLAN Membership.

But Dom0 does not need to deal with new device specific controls.◦ All required logic can be implemented in device specific drivers

and/or network management daemons.

Device Domain bridge is created for each VLAN.◦ Each does simple Bridging within a VLAN (802.1D).

Script (vif-bridge) is run whenever a VIF is added for a DomU.◦ Controls which Bridge it is attached to.◦ May specific the MAC Address

Default is random within a reserved OUI.Device Domain’s OS may provide other filtering/controls◦ Such as netfilter.

Many methods possible◦ As though 802.1Q Bridge per external port.◦ Static defaults applied to be unmanaged switch

All VNIC MAC Addresses are Manufacturer supplied.◦ Privileged operations via the Native Driver

Enabled for DomD or stand-alone Native Drivers.◦ Combinations of the above.

Existing vif-bridge script could easily configure the vsport matching the VNIC for a directly assigned VIF.◦ It already has MAC Address and any VLAN ID.◦ Suggested naming convention: use PCI Function number to name

the Backend instance. Simplifies pairing with direct device.

Multi-function NIC is unlikely to fully support all netfilter rules in hardware.When considering Direct Assignment:◦ Determine which netfilter rules are implemented by the

Multifunction NICs frame forwarding services.◦ Determine if the remaining netfilter rules can be trusted to DomU.◦ If there are filters that the hardware cannot implement, and

cannot be trusted to DomU, then don’t do the direct assignment.Direct Assignment complements frontend/backend.It is not a replacement.

Same-to-same migration only requires checkpoint/restore of any Device state via VF Driver.◦ Once state is checkpointed in VM memory, the

Hypervisor knows how to migrate the VM.Many services do not require migration◦ Each VM implements one instance of a distributed

Service. Persistence is a shared responsibility.◦ Most Web servers fall in this category.GOS probably already provides failover between dissimilar devices through bonding drivers.

Directly Assigned Devices can be migrated using existing services:◦ GOS Device Bonding / Failover.

Including netfront/netback in the team enables migrations between platforms with different hardware.

◦ GOS support of PCI device insertion/removal.Including check-pointing of any stateful data in host memory.

GOS already knows how to failover between two NICs.◦ Same-to-same migration requires same GOS

services whether failing over between two physical NICs or as a result of VM migration.

XEN already provides a universally available NIC via para-virtualization.◦ So a same-to-same Migration is always possible.

Requirement: device must be able to checkpoint any per-client stateful image in the client’s memory space.Device is told when to checkpoint any Guest-specific stateful information in the Guest memory image.Migrating Guest check-pointed memory image is a known problem that is already solved.Device driver on new host is told to restore from check-pointed memory image.◦ Checkpointed image should be devoid of any

absolute (non-VF relative) references.

Not all platforms have the same direct-access NICs, but same-to-same migration can be used anyway.Method A: post-migration makes right◦ Just do a Same-to-Same Migration anyway.◦ It will work

Of course because the actual device is missing on the new platform the re-activated instance will fail.Invoking existing device failover logic within the Guest.

◦ Possible Enhancement:Provide PCI Device removal event immediately on migration.

Method B: migrate Same-to-same via netfront.◦ Fail the Directly Assigned device.◦ GOS will failover to the Frontend device.◦ Migrate same-to-same to the new target platform.

Which always can support netfront.◦ Enable the appropriate Directly Assigned device on the new

platform.◦ GOS is informed on newly inserted PCI Function.◦ GOS will failover to the preferred device as though it were being

restored to service.

View of X3100 functions in Dom0◦ >>lspci –d 17d5:5833◦ 05:00.0 Ethernet controller: Neterion Inc.: Unknown device 5833 (rev 01)◦ --◦ 05:00.7 Ethernet controller: Neterion Inc.: Unknown device 5833 (rev 01)Export a function to DomU, so vxge driver can be loaded:◦ >>echo -n 0000:05:00.1 > /sys/bus/pci/drivers/vxge/unbind◦ >>echo -n 0000:05:00.1 > /sys/bus/pci/drivers/pciback/new_slot◦ >>echo -n 0000:05:00.1 > /sys/bus/pci/drivers/pciback/bind◦ >>xm pci-attach 1 0000:05:00.1Configure Bonding interface in DomU with active-backup policy and arplink monitoring, so the delegated interface (say eth1) can be enslaved to the bonding interface.◦ >>modprobe bonding mode=1 arp_interval=100 arp_ip_target=17.10.10.1

◦ >>ifconfig bond0 17.10.10.2 up◦ >>echo +eth1 > /sys/class/net/bond0/bonding/slaves◦ >>echo eth1 > /sys/class/net/bond0/bonding/primary

Create a virtual interface to enslave in DomU as a backup, and remove the delegated interface from DomU to transfer traffic to the virtual interface.◦ >>xm network-attach 1 bridge=xenbr1 ◦ >>echo +eth2 > /sys/class/net/bond0/bonding/slaves◦ >>echo –eth1 > /sys/class/net/bond0/bonding/slaves◦ >>ifconfig eth1 down◦ >>rmmod vxgeDetach pci function from DomU and migrate the DomU to the destination Xen machine (say IP is 172.10.9.7).◦ >>xm pci-detach 1 0000:05.00.1◦ >>xm migrate --live 1 172.10.9.7At this point, network traffic runs on the virtual interface on the destination machine

Delegate function on the destination machine◦ >>xm pci-attach 1 0000:02:00.1Enslave the direct interface◦ echo +eth3 > /sys/class/net/bond0/bonding/slaves◦ echo eth3 > /sys/class/net/bond0/bonding/primary◦ echo -eth2 > /sys/class/net/bond0/bonding/slavesRemove virtual backup in DomU◦ >>xm network-list 1◦ >>Idx BE MAC Addr . handle state evt-ch tx-/rx-ring-ref BE-path◦ 0 0 00:16:3e:45:de:53 0 4 8 768 /769 /local/domain/0/backend/vif/1/0◦ 2 0 00:16:3e:61:91:0a 2 4 9 1281 /1280 /local/domain/0/backend/vif/1/2◦ >>xm network-detach 1 2