Data centers have become an important computing platform
today, and the data-intensive workloads are driving modern
data centers to use higher-bandwidth networks, such as 10
Gigabit Ethernet (10GbE). However, traditional delay and
queue based traffic (congestion) control mechanisms for TCP
scale poorly with the high-bandwidth, low-latency data center
networks, and leaves 60% bandwidth (6Gbps in 10Gbps
link capacity) unused in typical data center workloads. This
problem severely constrains the performance of the data center
applications running on it.
To eliminate this inefficacy, we propose to couple congestion
among flows that span racks of physical hosts. Specifically,
we design the Wireless Sensing TCP (WSTCP) - a
combined low-end wireless and high-end wired network system
that combines the real-time coarse-grained traffic information
with packet loss to effectively coordinate the transmission
rate of multiple end systems, and implement it on
250Kbps Telos B wireless sensor nodes and 10Gbps Ethernet.
We evaluate WSTCP using various benchmarks and
Hadoop trace experiments on a 10GbE research testbed that
consists of 27 nodes. Experiment results demonstrate that
WSTCP improves the average throughput of TCP flows by
31-69%.
Check out our WSTCP paper draft submitted to ACM MobiSys 2014 for review.
