NETWORK CENTRIC PERFORMANCE IMPROVEMENT FOR LIVE VM

MIGRATION

ROBAYET NASIM, ANDREAS KASSLER

KARLSTAD UNIVERSITY ROBAYET.NASIM@KAU.SE

IEEE CLOUD 2015, NEW YORK, USA, JUNE 2015.

TRAFFIC COEXISTENCE IN DATACENTER NETWORKS

Aggregation Switches

K Pods with K Switches

each

K=4

Racks of servers

vm vm vm vm

vm vm vm vm

VM-VM traffic

Migration traffic

Migration traffic

Migrate communicating VMs towards the same server to maximize performance

Latency depends on – VM Placement

– Other VMs Background traffic

– Migration Traffic

– Queuing strategies of Datacenter switches

Implemented Public Transport App based on Pub/Sub – Evaluate Latency in local

Openstack testbed

– Different queuing strategies

HOW TO CONTROL VM TO VM LATENCY?

Evaluate the impact of utilizing multiple paths using MPTCP and different queuing strategies for live VM migration traffic

Architecture for Flexible Prioritization of: – VM-to-VM traffic

– Live VM Migration Traffic

Based on OpenStack Neutron and OpenDaylight SDN

Uses OpenVSwitch with different queuing strategies configurable via ODL – HTB, CODEL, FQ_CODEL (fair queuing controlled latency, not support by

traditional datacenter switches) • See: QoS Enabled WiFi MAC Layer Processing as an Example of a NFV Service, Jonathan Vestin, Andreas

Kassler, to appear in: Proceedings of the IEEE NetSoft2015, April 13-17 2015, London, UK.

MAIN CONTRIBUTIONS

LIVE VM MIGRATION

QUEUING STRATEGIES

FQ_CODEL

HTB

ARCHITECTURE

Datacenter Switches

Flexible priorization of Migration versus VM traffic

Enable MPTCP

Enable MPTCP

Live Vm Migration traffic: KVM QEMU, libvirt port range 49152 …49216,16509

Queue configuration for port 4 (c06df192-b6... ) on OVS using ODL Rest conf API

Configuring the Live VM Migration traffic Queue connected to QoS uuid 109b55a6-1e... :

OF Rule

EXAMPLE QUEUE CONFIGURATION AND OF RULES

{ "parent_uuid": "c06df192-b6bc-4f...", "row": { "QoS": { "other_config": [ "map", [[ "max-rate", "1000000000" ]]], "type": "linux-htb" }}}

{ "parent_uuid": “109b55a6-1ea6-42...", "row": { ”Queue": { "other_config": [ "map", [[ ”priority", ”1" ]]], }}}

in_port=3,tp_src=16509,actions:enqueue:4:1

EXPERIMENTAL SETUP

OpenStack Controller OpenStack Compute1 OpenStack Compute2

Gigabit Switch Open vSwitch

Traffic Generator Traffic Generator

Bottleneck link

OpenStack Internal network

OpenStack VM Network

Cross Traffic

Traffic Generator Traffic Generator

OpenDaylight Controller

Neutron Interface

ML2 Driver

Open vSwitch Configure

VM-to-VM path, em2

VM-to-VM path, em2

Management path, em1

Management path, em1

Extended to support HTB, CoDEL, Fq_CoDel,

etc.

Extended to support HTB, CoDEL, Fq_CoDel,

etc.

Extended to support HTB, CoDEL, Fq_CoDel,

etc.

Extended to support HTB, CoDEL, Fq_CoDel,

etc.

KVM-QEMU uses port range 49152-49216

KVM-QEMU uses port range 49152-49216

Libvirt uses 16509 Libvirt uses 16509

Openflow rules Openflow rules

Case 1: Unloaded VM

Case 2: VM with Memory Intensive and Storage-heavy Applications

Case 3: Impact of background load and queuing strategies.

Case 4: Real-time ITS application scenario

Different flavors of VMs in OpenStack

EVALUATION SCENARIOS

CASE 1: UNLOADED VM

Using MPTCP for Migration traffic reduction in Downtime

For medium VM (3 GB memory, 3.3 GB files on storage)

For large VM (6 GB memory, 5.8 GB files on storage)

For X-large VM (12 GB memory, 7.8 GB files on storage)

CASE 2: STRESS TEST

Using MPTCP for Migration traffic significant reduction in Downtime

VM UNDER DISK AND MEMORY LOAD TOGETHER

Using MPTCP for Migration traffic significant reduction in Downtime

CASE4: DIFFERENT NETWORK LOAD ON THE NETWORK PATHS

Management path, em1

Management path, em1

VM-to-VM path, em2

VM-to-VM path, em2

UDP flows at 500/800 Mbps

Netperf wrapper (8 tcp flows in both directions)

CASE 4: IMPACT OF DIFFERENT NETWORK LOADS AND QUEUEING STRATEGIES

Two HTB queues, 200 Mbps for VM-to-VM traffic, 800 Mbps for migration traffic

Two HTB queues, 200 Mbps for VM-to-VM traffic, 800 Mbps for migration traffic

OpenDaylight Rest API: Create flow rules to enqueue all management traffic and VM-to-VM traffic to the proper queues.

Using MPTCP for Migration traffic significant reduction in Downtime Prioritization of Mgmt traffic: better performance

CASE 4: REAL-TIME PUBLIC TRANSIT TRACKING APPLICATION SCENARIO

FQ_CODEL helps to keep latency of VM to VM traffic controlled

CASE 4: REAL-TIME PUBLIC TRANSIT TRACKING APPLICATION SCENARIO

MPTCP for migration + FQ_CODEL: shortest application latency

Conclusions – Combined different mechanisms at network and transport layer to

improve the performance of live VM migrations.

– Flexible architecture with SDN and Opendaylight.

– MPTCP for migration traffic reduces VM downtime significantly.

– FQ_CODEL allows to control VM-to-VM latency even under congestion.

Future work – Evaluation with larger topologies and diverse link

capabilities.

– Energy efficient traffic routing of migration and VM-to-VM traffic.

CONCLUSIONS AND FUTURE WORK

Thank you for your attention!

Q/A