THESIS
2017
xix, 130 pages : illustrations ; 30 cm
Abstract
Nowadays, data centers are being built around the world to provide infrastructure
support for big data and cloud computing. Many data center applications such as web
search, recommendation systems, social networking, etc., have very demanding latency
requirements. Even a small delay due to network congestion can directly affect application
performance and degrade user experience. Therefore, handling the network congestion in
data centers is critical.
This thesis focuses on congestion control mechanisms for data center networks. Specifically, we make the following three key contributions.
First, we present PIAS, a flow scheduling mechanism to minimize the average
flow completion time. PIAS differs from most prior solutions in that it: 1) does not assume prior
knowledge of flow...[
Read more ]
Nowadays, data centers are being built around the world to provide infrastructure
support for big data and cloud computing. Many data center applications such as web
search, recommendation systems, social networking, etc., have very demanding latency
requirements. Even a small delay due to network congestion can directly affect application
performance and degrade user experience. Therefore, handling the network congestion in
data centers is critical.
This thesis focuses on congestion control mechanisms for data center networks. Specifically, we make the following three key contributions.
First, we present PIAS, a flow scheduling mechanism to minimize the average
flow completion time. PIAS differs from most prior solutions in that it: 1) does not assume prior
knowledge of flow size information; and 2) can be readily implemented with commodity
switch hardware and legacy network stacks.
Second, we reveal that existing Explicit Congestion Notification (ECN) marking
schemes suffer from severe performance impairments in multi-service multi-queue data
centers. Then we propose MQ-ECN, a new ECN marking scheme for multi-queue with
round-robin scheduling, which is widely used in production data centers. Moreover, driven
by recent progress in programmable schedulers, we further design TCN that enables ECN
for multi-queue with arbitrary scheduling.
Third, we observe that existing transport solutions suffer from either excessive packet
losses or serious throughput degradation in high-speed extremely shallow-buffered data center networks. Then we propose BCC, a simple yet effective solution with only one more
ECN configuration at the switch. BCC maintains low packet loss rate persistently while
keeping high throughput until the buffer becomes insufficient.
Post a Comment