Slide - csie.org
Download
Report
Transcript Slide - csie.org
IMPROVING DATACENTER
PERFORMANCE AND
ROBUSTNESS WITH
MULTIPATH TCP
Costin Raiciu†, Sebastien Barre‡, Christopher Pluntke†,
Adam Greenhalgh†, Damon Wischik†, Mark Handley†
Before Start
我想要先用我的話說一次這篇Paper想做的事情:
這邊Paper的重點在於,single path用在TCP上面
的時代已經過去了,multipath TCP才是主流。
Multipath TCP又主要有三種topologies: fattree,
VL2, Bcube。
這篇paper又分析了在不同情況下multipath確實
比single path優秀的地方。還有一些影響
performance的因素的分析。
最後是有沒有其他可能再發展的topology
OUTLINE
Introduction
Data center networking
Topology.
Routing.
Path Selection.
Congestion Control.
Multipath TCP in summary
MPTCP in data center
Examples of Benefits
Analysis
Evolving topologies with MPTCP
Influence of Topology
Number of Subflows
Influence of the Traffic Matrix
Analysis
Discussion about DHFT
Summary
Introduction
Traditional single path TCP has many significant
problems, ill-suited for such network. (intra-data
center traffic.)
Ex. Bottleneck, can’t scale up, congestion.
We propose using Multipath TCP as a replacement
for TCP in such data centers.
Introduction
Traditional data center topology.
…
Introduction
4 components to a data center architecture:
1. physical topology
2. routing
3. path selection
4. congestion control of traffic on the selected path.
OUTLINE
Introduction
Data center networking
Topology.
Routing.
Path Selection.
Congestion Control.
Multipath TCP in summary
MPTCP in data center
Examples of Benefits
Analysis
Evolving topologies with MPTCP
Influence of Topology
Number of Subflows
Influence of the Traffic Matrix
Analysis
Discussion about DHFT
Summary
Data center networking---Topology
Traditional data center topology.
…
Data center networking---Topology
That has many problems and can be improved a lot
like we have discussed before.
So, fat-tree is one kind of solutions.
Fat Tree Topology [Fares et al., 2008; Clos, 1953]
K=4
Aggregation
Switches
K Pods with
K Switches
each
Racks of
servers
Data center networking---Topology
And we also have VL2 and Bcube these two kinds
of solutions.
BCube Topology [Guo et al, 2009]
BCube (4,1)
Data center networking---Topology
VL2 Topology [Greenberg et al, 2009, Clos
topology]
10Gbps
10Gbps
…
20 hosts
Data center networking---Routing
We cannot expect the host itself to know which of
these paths is the least loaded, so the routing
system must spread traffic across these paths.
Use randomized load balancing, where each flow
is assigned a random path from the set of possible
paths.
Data center networking--Path Selection
randomized load balancing有他一定的問題,就
是有一定的機率某些path會有超大流量、同時
會有一些path沒什麼(極低或零)流量。
The use of a centralized flow scheduler has been
proposed.
Multipath TCP: Congestion Control
[NSDI, 2011]
Data center networking--Congestion Control
Our hypothesis is that given sufficiently many
randomly chosen paths, MPTCP will find at least one
good unloaded path, and move most of its traffic
that way. In so doing it will relieve congestion on
links that got more than their fair share of ECMP
balanced flows.
Multipath TCP in summary
By making the window increase depend on the total
window size, subflows that have large windows
increase faster than subflows with small windows.
OUTLINE
Introduction
Data center networking
Topology.
Routing.
Path Selection.
Congestion Control.
Multipath TCP in summary
MPTCP in data center
Examples of Benefits
Analysis
Influence of Topology
Number of Subflows
Influence of the Traffic Matrix
Evolving topologies with MPTCP
Analysis
Discussion about DHFT
MPTCP in data center
Three main benefits:
1. Better aggregate throughput
2. Better fairness
3. Better robustness
MPTCP in data center
Depend on 4 requirements:
1. The congestion control scheme used.
2. The physical topology.
3. The traffic matrix generated by the applications.
4. The level of load in the network.
Examples of Benefits
Throughput:
Using MPTCP in a FatTree network, we can see the
more subflows we have, the more throughput we
have.
The reason is simple, n flows share one link, each
flow need to achieves 1/n capacity.
Examples of Benefits
Examples of Benefits
Fairness:
Every host’s throughput is shown ranked in order of
increasing throughput. Is is clear that not only did
the utilization improve with MPTCP, but also the
fairness improved.
Examples of Benefits
這是一個累積百分比的圖,可以看出single TCP
有些path perform 真的很差(很閒的意思),而
MPTCP則大幅改善此情況。
Analysis
Then, we have questions:
How many subflows are needed?
How does the topology affect results?
How does the traffic matrix affect results?
Analysis
We need to view in these aspects to answer the
question:
Influence of Topology
Number of Subflows
Influence of the Traffic Matrix
Influence of Topology
We compare VL2 and Bcube.
Influence of Topology
Why Bcube is relative low?
I guess, just guess, is because in Bcube, pathes
between two hosts are not in same length which
lead to worse flows management.
Number of Subflows
We were surprised that eight subflows were
needed to achieve 90% throughput.
Why eight, and on what does this depend?
Number of Subflows
Results:
Influence of the Traffic Matrix
OUTLINE
Introduction
Data center networking
Topology.
Routing.
Path Selection.
Congestion Control.
Multipath TCP in summary
MPTCP in data center
Examples of Benefits
Analysis
Influence of Topology
Number of Subflows
Influence of the Traffic Matrix
Evolving topologies with MPTCP
Analysis
Discussion about DHFT
Summary
Evolving topologies with MPTCP
In single homed topologies:
Hosts links are often bottlenecked.
ToR switches failure also lead to crash tens of hosts.
Evolving topologies with MPTCP
以FatTree為例子
Evolving topologies with MPTCP
So, how do we improved?
There are two solutions: perfect switch and dualhomed FatTree.
Evolving topologies with MPTCP
Perfect switch:
Serves as a good control experiment, giving an
upper bound on what any network core might
provide using single links to the hosts.
Evolving topologies with MPTCP
Dual-homed FatTree (DHFT):
is to the host and four connect the links between the
two layers of switches.
If we remove one port per host from the core and
use it to connect the second interface on each
server, the network requires the same number of
switch ports.
Analysis
QUESTION:
Will perfect switch really improve performance?
Analysis
Testing throughput over MPTCP, perfect switch TCP,
normal TCP.
Analysis
QUESTION:
Will perfect switch really improve performance?
ANSWER:
Yes, it is.
Analysis
QUESTION:
Dual-homed FatTree really work?
Does it have any problem?When may troubles
occur?
Analysis
Testing relative throughput over MPTCP, TCP DHFT,
TCP perfect switch.
Analysis
QUESTION:
Dual-homed FatTree really work?
Does it have any problem?When may troubles
occur?
ANSWER:
Yes, it works only when load is low. When load is
high, throughput will go down near to TCP.
Discussion about DHFT
DHFT’s worst case performance is 75% and best case is
around 200%.
在流量不穩定的情況下,DHFT + MPTCP有可能會有
較差的表現產生。
Beyond performance, DHFT improves robustness: any
lowerpod switch failure does not cut-off an entire rack
of servers.
DHFT is not optimal by any measure, but it shows that
we can create topologies with better performance if
we assume MPTCP is the transport protocol.
OUTLINE
Introduction
Data center networking
Topology.
Routing.
Path Selection.
Congestion Control.
Multipath TCP in summary
MPTCP in data center
Examples of Benefits
Analysis
Evolving topologies with MPTCP
Influence of Topology
Number of Subflows
Influence of the Traffic Matrix
Analysis
Discussion about DHFT
Summary
Summary
“One flow, one path” thinking has constrained
datacenter design
Collisions,
unfairness, limited utilization
Multipath transport enables resource pooling in
datacenter networks:
Improves
throughput
Improves fairness
Improves robustness
Reference
conferences.sigcomm.org/sigcomm/2011/slides/s2
66.ppt
Q & (A)