Approved for Public Release
This work was supported in part by the EPSRC INTERNET Project EP/H040536/1, the Defense Advanced Research Projects Agency (DARPA) and the Air Force Research Laboratory (AFRL), under contract FA8750-11-C-0249. The views, opinions, and/or findings contained in this article/presentation are those of the author/presenter and should not be interpreted as representing the official views or policies, either expressed or implied, of the Defense Advanced Research Projects Agency or the Department of Defense.
R2D2: Bufferless, Switchless Data Center Networks Using Commodity Ethernet Hardware
Matthew P. Grosvenor Malte Schwarzkopf Robert N. M. Watson Andrew W. Moore
r2d2@srg
http://goo.gl/U3rL85
Switch Scheduler
Statistical Multiplexing Switch
STOP
Host's queues Switch queues
Unpredictable outputOn the wire
Time = 0.8µs to >1900µs
Network Scheduler Passive Cross Point (PCP)
Host's queues On the wire
Time = Fixed (0.8µs to <3.5µs)
Tim
e =
Pred
icta
ble
Predictable outputNo queues
Statistical multiplexing uses queueing, which causes unpredictable long latency tails
Explicit network scheduling allows us to deterministically multiplex without long tails
PCP PCP
PCP
Gateway
Passive Cross Point
The Resilient Realtime Data Distributor (R2D2) network architecture
Network Scheduler
Router
Arbiter
Gateway
Latency CDF: Merging 500Mb/s + 5Gb/s flows through various switches [256B packets]
DownlinkUplink
Latency CDF: Packets traversing an R2D2 network using a prototype software gateway with varying packet size.
Come talk to us about bandwidth, scheduling policies and applications for R2D2!
Ideally implemented in hardware. Our prototype achieves good results in software only.
Can be implemented on a NetFPGA or using a subset of a regular switch.
Latency CDF: Merging 2x 10Gb/s flows through various switches [256B packets]
Latency CDF: NetFPGA 10G PCP at varying packet sizes.
Uncontended latency vs packet size, measured through several PCP prototypes.
Late
ncy
[µs]