S t a n f o r d U n i v e r s i t y

Hypervisor based DataCenter wide Network Provisioning Vimalkumar Jeyakumar, Mohammad Alizadeh, Balaji Prabhakar, David Mazières (Stanford University)

Changhoon Kim, Albert Greenberg (Windows Azure)

Today’s multi-tenant cloud

§  Unpredictable performance: “Noisy neighbour” §  Netflix: AWS networking has more variable latency… Ø Re-architect applications

§  Can provision CPU, memory, disk: familiar units §  Network?? Ø Growing bandwidth, stringent latency reqts.

Provisioning Network for Cloud Apps in Gbps §  Simple model: per-VM VNIC capacity ü  Absolute units: Mbps ü Minimum guarantee: high utilisation ü  VMs connected to single switch ü  Full bisection bandwidth ü  Predictable performance

TOR

Agg

Core

Big Switch (DC Network)

EyeQ: Hypervisor Shim

Internet

~100Gbps

~1Pbps

VMs

!"#!"#

!"#$"#

!"$%&'()#

!"# !"#

*+#,-'%./0#

!"$%&'()#123456-#73'-#

8&9&'-/#

7:#(.94.;-;'#2-'-('<#

=:#(.94.;-;'#/-3('<#

+.;>-<5.;#*-'-('./#

=:#43(0-'<# 7:#43(0-'<#

?;'/?<'-2# ?;'/?<'-2#

%&'()*(+&'#,*-#.*&)/*0#

1234-5).67234-5).68##

!"#

@AA"B4<#CAA"B4<#

§  Address contention and sharing ü  At end hosts ü  At Internet uplink

§  Trends: high bandwidth interconnects

§  EyeQ: Eye for Quality o  Enforce at TX: adaptive rate limiters o  Detect at RX: congestion signaller

§  Distributed mechanism o  End to end control loop o  Fine timescales reactions: few 100µs

Sample Packets diverted to end host

!"#$#%&'()&*+,%-$.&/0+%12&

*34& *35& *3#&*36& !"

!"#$#%&7()&*+,%-$.&/0+%12&

*34& *35& *3+&*36& !"

!"#$%&'()*)#+',-.'#)$/')$/-,01'##)%2)'3,/)+-+')

1.6GHz 4GB 1TB “High”?!

This work is supported by NSF award CNS-1040190 (FIA: NEBULA)