1. 2018/7/4
QoS-aware Multipathing in datacenters using effective bandwidth estimation and SDN
Author: Runxin Wang, Simone Mangiante, Alan Davy, Lei Shi, Brendan Jennings
Presenter: Xiaomin Zheng
Date:2017/02/22
2016 International Federation for Information Processing(IFIP)
Department of Computer Science and Information Engineering
National Cheng Kung University, Taiwan R.O.C.
CSIE CIAL Lab 1

2. 2018/7/4
Introduction
Without a dedicated flow control mechanism, different sizes of traffic flows are statically allocated to links without sensing the current link utilization, which can result in transient network congestion that impacts on the latency experienced by users of the hosted applications.
Existing works aim to realize load balancing by efficiently scheduling flows based on their sizes and link utilization.
Many applications’ performance is bounded with specific QoS targets, so load balancing may not directly address their QoS requirements.
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

3. 2018/7/4
Introduction
We present a latency-aware flow scheduling system that schedules flows based on a tenant application’s QoS target and empirical estimations of the effective bandwidth required to meet these QoS targets.
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

4. 2018/7/4
Introduction
Equal-cost multi-path (ECMP) is the common protocol used in datacenter networks to map flows into multi-equal paths. It’s deficiency in the presence of larger flows, which often leads to link congestion onset.
On the other hand, most proposed flow scheduling techniques ignore the individual QoS required by each VM—they only ensure that, overall, the flows are evenly allocated to the given links, but they do not directly address the network resources needed by each VM.
Equal-cost multi-path (ECMP) is a routing technique for routing packets along multiple paths of equal cost.
每一封包多路徑路由的負載平衡通常不適用因為大輻變化的延遲、數據包重新排序
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

5. 2018/7/4
Introduction
The issue can be usually found in the resource allocation approaches that are based on end host measurement and end host congestion feedback mechanisms
we leverage the Software Defined Networking (SDN) paradigm to develop a flow scheduler, termed the MAPLE-Scheduler, which monitors the network changes in switches and dynamically reschedules flows to meet the QoS requirements, while balancing the link utilization across links.
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

6. 2018/7/4
Related work
DevoFlow achieves a scalable flow management: it keeps flows in the data plane as much as possible to avoid overheads in the controller, but it maintains enough visibility over network flows to provide effective aggregated flow statistics.
It uses a trigger-based flow statistic gathering mechanism which inspired us to design an efficient link monitoring component.
All OpenFlow solutions can be built on top of DevoFlow; however, DevoFlow enables scalable implementation of these solutions by reducing the number of flows that interact with the control-plane.
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

7. The architecture of MAPLE system
2018/7/4
The architecture of MAPLE system
MAPLE is a network-aware VM ensemble placement system that uses empirical estimations of the effective bandwidth required between servers to ensure that QoS violations.
The MAPLE Controller processes VM placement requests from tenants and instructs the Cluster Manager, which in turn configures VMs on the selected servers.
EB agent: They are responsible for collecting traces of traffic emanating from the server using utilities such as tshark. 隔个几秒记下测量的effective bandwidth， controller要的时候才会向EB Agent拿。
It requests the Residual Bandwidth Estimator to provide it with a set of candidate servers for the placement and the estimates of residual bandwidth available on the egress links of those servers.
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

8. The architecture of MAPLE system
National Cheng Kung University CSIE Computer & Internet Architecture Lab

9. The logic architecture of MAPLE-Scheduler and MAPLE system
2018/7/4
The logic architecture of MAPLE-Scheduler and MAPLE system
Distributed monitoring agent
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

10. Effective Bandwidth In Multipath Topologies
2018/7/4
Effective Bandwidth In Multipath Topologies
EB estimation needs to be revisited in order to take into account that it is practically and computationally hard to measure and run the EB estimation algorithm at every single link in the network.
We logically divide the network in two parts—the edge and the core—and accordingly use two different approaches to conduct EB estimation.
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

11. Effective Bandwidth In Multipath Topologies
At the edge of the network, where top-of-rack switches (ToR) are, we apply the precise EB estimation method for every uplink port connecting the ToR switch to the core.
Since the ToR switch aggregates traffic coming from the host and the VMs in the same rack, it is practically feasible to make the link selection (flow scheduling) decision
National Cheng Kung University CSIE Computer & Internet Architecture Lab

12. Effective Bandwidth In Multipath Topologies
The core of the datacenter network, we employ the approach of Effective Bandwidth Coefficient (EBC) to estimate the EB for the observed links.
EBC captures the relation between estimated EB and its corresponding mean throughput of a link regarding a specified QoS target.
Reff,i represent the EB of a given link i, meani is the current mean throughput of the link,
National Cheng Kung University CSIE Computer & Internet Architecture Lab

13. 2018/7/4
MAPLE-Scheduler
The MAPLE-Scheduler is a flow scheduling system that monitors flows’ QoS performance, and dynamically reschedules flows, aiming to maintain their QoS needs.
The key components of MAPLE-Scheduler
A centralized controller built on SDN technologies
A flow scheduling algorithm making schedule decisions
Distributed monitoring agents residing on servers and edge switches
这个scheduler可以支援各种topology，但是switch必须可以支援openflow
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

14. 2018/7/4
MAPLE-Scheduler
It schedules flows based on the availability of sufficient residual bandwidth, calculated based on estimations of EB.
effective bandwidths are estimated based on captured traces
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

15. 2018/7/4
MAPLE-Scheduler
In this work we implemented a trigger based push method for gathering flow statistics, with the aim to reduce the work loads and enable the controller to be timely responsive to state changes in switches.
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

16. 2018/7/4
MAPLE Flow Scheduling
The MAPLE-Scheduler only reschedules large flows with the aim to maintain QoS targets.
It reschedules a flow based on two conditions:
if it will not exceed the selected link’s residual bandwidth, based on the EB estimations;
that max-min fairness is met max-min fairness is applied to achieve fair loads of traffic distributed across a multipath network.
a flow is large flow if it consumes more than α = 10% of its hosting link’s capacity.
Max-min: maximizing the residual bandwidth of the link that has minimum residual bandwidth
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

17. MAPLE Flow Scheduling
National Cheng Kung University CSIE Computer & Internet Architecture Lab

18. 2018/7/4
Evaluation
Switches within the pods are Open vSwitch instances with 1 Gbps upstream and downstream capacity.
The emulated datacenter has a single root switch connecting 4 physical servers. One server is used as the management server; the rest servers hosting 210 VMs that are connected by 2 levels of vSwitches.
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

19. 2018/7/4
Evaluation
To emulate bulk data transfers within the network, we used the following data applications:
Hadoop application;
A data serving application that combines YCSB (data client) and Cassandra (data server);
A custom data backup application, created using the SCP utility.
To simulate large flows that effectively change the QoS performance, we run iperf on a dedicated VM to some randomly selected VMs.
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

20. Evaluation The Analysis of QoS violations
2018/7/4
Evaluation
The Analysis of QoS violations
National Cheng Kung University CSIE Computer & Internet Architecture Lab
CSIE CIAL Lab

21. Evaluation Analysis of Throughput Performance and Link Utilization
QoS target (0.02s, 0.1) QoS target (0.04s, 0.05)
National Cheng Kung University CSIE Computer & Internet Architecture Lab

22. Evaluation
The gap between two lines on the range of link utilization between 0.2 to 0.4 that indicates MAPLE Scheduler achieve lighter loads across its links, comparing to ECMP
National Cheng Kung University CSIE Computer & Internet Architecture Lab

23. Conclusion
we designed MAPLE Scheduler built on SDN controller, which can effectively schedule flows to enforce QoS guarantee in multipathing datacenters.
National Cheng Kung University CSIE Computer & Internet Architecture Lab