Consistent Resource Scheduling and QoS management

Applied Research Center
for Computer Networks
Moscow State University
Ruslan Smelyanskiy
ARCCN Director, Professor at Moscow State University
Consistent Resource Scheduling and
QoS management

Content
 Brief ARCCN introduction and main
research direction overview
 Two research problems examples:
 The Consistent Resource Scheduling in DC
 FDM-TCP: Flow DeMultiplexing TCP for QoS
management
Huawei Shenzhen R.Smelyanskiy 226.05.2015

ARCCN
Research
Ideas
Education
Adoption
Cooperation
Development
Integration
Applied Research Center for Computer Networks (ARCCN) is a Russian
non-profit organization initially funded by Skolkovo intended to:
 world-class competency center for Computer Networking in Russia;
 collaboration between Russian and international research, scientific,
educational and commercial experts and institutions;
 demonstrate innovations in Networking for National Industry;
 promote commercialization of a new networking technologies and
services.
ARCCN is the only R&D competency center for SDN&NFV
in Russia today
Partners: Skolkovo, The Ministry of Education and Science of Russia,
Rostelecom, Rostech, Sberbank, EMC, Intel, etc.

Project Ecosystem
7%
20%
12%54%
7%
Ph.D.
Assos. Prof
Postgraduate
Engineeres
ARCCN Team

Key R&D directions:
 Distributed SDN Controller and Applications for it;
 Safety and Security in SDN;
 QoS management in SDN;
 NFV platform and VN functions;
 Architecture for Open Flow-switch;
 WAN operation prototyping and analysis;
 Self Organizing Cloud for Carriers (Cloud Conductor);
 Data Center Resource Management;
 Forwarding policy validation and troubleshooting automation;
 Software for integration with legacy transport on the Internet.

Consistent Resource Scheduling in DC

Scheduler determines a physical
resource for each element of a
tenant

Data center resources model
Physical resource model: ,where
• Р - set of computational nodes:
– vh(p) - number of CPU cores on node p
– qh(p) - amount of RAM on node p
• М - set of data storages:
– uh(m) – size of data storage m
– type(m) - m data storage type
• К - set of network switches:
– - bandwidth of the switch k
• L - set of network channels:
– rh(l) – bandwidth of the channel l
),( LKMPH 
( )h k

Tenant model
Tenant model:
• W - set of virtual machines:
– v(w) – requested number of cores for VM w
– q(w) – requested amount of RAM for VM w
• S - set of storages:
– u(s) – size of storage s
– type(s) – s storage type
• E - set of virtual channels:
– r(e) – bandwidth of the virtual channel e
Tenant types:
• Loosely coupled:
• Tightly coupled:
),( ESWG 
)]),}{,}[({ 11  ESW K
jj
N
ii
)}]}{,}{{(),}{,}[({ 1
'
1
'
11
K
jj
N
ii
K
jj
N
ii SSWWESW  

Tenant Mapping on DC Resources
where
Mapping A is correct if the following constraints are fulfilled:
a)
b)
c)
d)
)()(),()( pqhwqpvhwv
pp WwWw
  
: { , , { , }},A G H W P S M E K L    
)()( lrher
lEe

)()( kher
kEe

),()( muhsu
mSs

)()(: mtypestypeSs m 
, , ; , , , W S E G P M K L H

Replication and migration
Replication:
• Duplicates storage m to m’
• Channel to support consistency between m and m’:
• VM connects to m’
Migration:
• Moves virtual machine or storage to another physical resource
'
: ii AAR 
( ', , , , , , ); ; ; ',
1 1 1
m l k k l m k K l L m m M
n n i i
  

Huawei Shenzhen R.Smelyanskiy 11
'
: ii AAT 
26.05.2015

Tenant mapping problem
The problem:
• For every tenant Gi from set of tenants Z = {Gi} and
given DC resources Hres define correct mapping Ai
that set A = {Ai} covers the maximum number of
tenants from Z

Residual graph
Renew Z – 𝒁 → 𝒁′
Residual graph Hres:
• The following functions are recalculated: vh(p), qh(p), rh(l), τh(k), uh(m)


pWw
res wvpvhpvh )()()( ( ) ( ) ( )
p
res
w W
qh p qh p q w

  
( ) ( ) ( )
l
res
e E
rh l rh l r e

  
( ) ( ) ( )
k
res
e E
h k h k r e 

  
( ) ( ) ( )
m
res
s S
uh m uh m u s

  

Scheduling Round
 The scheduling algorithm runs at the begging of the
scheduling rounds
 On a scheduling round:
– Add new tenants to Z
– Delete the tenants with TTL=0
– Recalculate Residual graph as a new H

Scheduling algorithms
• A1: greedy and limited exhaustive search strategies
– maps tenants as a whole (minimal common subgraph
isomorphism algorithm)
– suitable when critical resource is a data center physical
network
• A2: greedy and limited exhaustive search strategies
– constructs mappings element by element (bin-packing
algorithm)
– suitable when critical resources are computational nodes or
data storages
• A3: Ant colony algorithm:
– universal, but greater computational complexity

Initial data for test
Test # DC model (graph H) Tenants models (graph G)
Test 1
450 computational nodes with
21000MB RAM and 16 cores
per node
450 data storages with
21000GB disk space per
storage
1350 virtual machines and 1350
storages total in 135 tenants.
Test 1 – loosely coupled tenants
Test 2 – tightly coupled tenants,
network load = 70%
Test 2
Test 3
500 computational nodes with
10000MB RAM and 16 cores
per node
500 data storages with
10000GB disk space per
storage
600 virtual machines and 600
storages per 100 tenants.
Test 3 – loosely coupled tenants
Test 4 – tightly coupled tenants,
network load = 70%
Test 4
Fattree topology with 5 TOR switches, 10 aggregation switches, 40 edge
switches
Huawei Shenzhen R.Smelyanskiy
1626.05.2015

Comparison with OpenStack
scheduling algorithms
Test #
Open stack CRM
FF RF
Test 1 50% 87.5% 100%
Test 2 0% 1% 100%
Test 3 100% 70% 100%
Test 4 0% 0% 100%
- FF – First Fit
- RF – Random Fit
- CRM – Consistent Resource Mapping

Conclusions
 Open Stack algorithms are:
– Ineffective for some cases of loosely coupled tenants
– Not applicable for tightly coupled tenants
 Consistent Resource Mapping algorithms:
– Effective as for loosely as for toughly coupled tenants
– Suitable for IaaS with SLAs

FDM-SDN: SDN with Flow
DeМultiplexing TCP
is a new way to manage the quality of service

Hosts do affect connection quality!
• TCP congestion avoidance algorithms:
– Goal: get connection with maximal bandwidth
– Strategy: cut-and-try to detect the maximum currently
available amount of resources to utilize all of them
– Primary heuristic: AIMD (pessimistic)
– Recovering action: decries CWND
– Parameter: congestion window size (CWND), timeout
– Primary modes: slow start and congestion avoidance
– Triggering criteria: duplicate ACKs&timeouts, threshold
• There is no way to control routes intersections in
traditional networks

FDM-SDN provides a new leverage
• Flow demultiplexing along several non-xing routes:
– Goal: get connection with required bandwidth
– Strategy: use cut-and-try to detect the amount of provided
resources and utilize all of them on several routes
– Primary heuristic: decries the number of routes
(optimistic)
– Parameter: cumulative bandwidth of all routes
– Primary mode: keep busy all routes
– Triggering criteria: bandwidth deficiency
– Recovering action: open new non-xing route

Flow DeMultiplexing Protocol
Standard socket API
Activity Monitor (Bandwidth Scarcity Detection)
TCP subflow
(extra options)
TCP subflow
(extra options)
TCP subflow
(extra options)
Application Layer
Transport Layer
Network Layer
FDM TCP (Packet Scheduling & Reordering)
Subflow Manager (Split Degree Adjustment)

Routing FDMP flows with SDN
Host A Host B
SDN Controller
Has MP_CAPABLE option?
Install new FDMP connection!
SYN
MP CAP
Key A
Set up new FDMP connection

Host A Host B
SDN Controller
Has MP_CAPABLE option?
Complete partial FDMP connection!
SYN, ACK
MP CAP
Key B
Set up new FDMP connection

Host A Host B
SDN Controller
Has MP_JOIN option?
Install new FDMP subflow for a
known connection!
SYN
MP JOIN
Token B
Set up new FDMP subflow

Host A Host B
SDN Controller
Subflow is not active any more!
Remove the path!
FDMP subflow manipulation
Actually, we store metadata and
allow some subflows to resume.
We use flow eviction to
remove this data.

Actually, we store metadata and
allow some subflows to resume.
We use flow eviction to
remove this data.
Host A Host B
SDN Controller
Get FDMP packet of a expired subflow!
Either reroute the remembered subflow,
or force hosts to close it
RSTRST
Close violet subflow!
Reschedule the packet to
red subflow!
Close violet subflow!
FDMP subflow manipulation

h2h1
lower = 50 Mbps
middle = 50 Mbps
upper = 50 Mbps
Experiments
FDMP in a hammock
TCP MP TCP FDMP
49 Mbps 48 Mbps 81 Mbps
Required h1 to h2 bandwidth: 80 Mbps

upper = 50 Mbps
h2h1
h4h3
lower = 50 Mbps
middle = 50 Mbps
Adjustment to congestion
TCP MP TCP FDMP
49 Mbps 48 Mbps 81 Mbps
Required h1 to h2 bandwidth: 80 Mbps

Conclusion
1. In context of bandwidth, the party of FDMP & our
routing application outperform both single-flow TCP
and MP TCP without regards to ECMP
2. FDMP improves efficiency of infrastructure:
It can increase utilization under the same load and
allows network to process more traffic
3. Under a heavy load FDMP will support faster
communication and/or transmit more data within
the same period of time

Challenges we are working on
 Use TCP-FDM to reduce delay and packet loss:
– Which packet scheduling algorithm to use?
– What routing algorithms are appropriate?
 Enforce resource allocation & QoS policies:
– How to avoid flow competition on congestion?
– Do we need to control request frequency?
– Is it efficient enough to rate-limit subflows?
– How to combine TCP-FDM and Diff Serv model?
 Develop FDMP as VNF within network:
– How to integrate TCP-FDM routing application?
– How to build an efficient TCP to TCP-FDM proxy?

http://arccn.ru/
+7 (495) 984 27 64
smel@arccn.ru
@ArccnNews

Consistent Resource Scheduling and QoS management

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Consistent Resource Scheduling and QoS management

Similar to Consistent Resource Scheduling and QoS management (20)

More from ARCCN

More from ARCCN (20)

Recently uploaded

Recently uploaded (20)

Consistent Resource Scheduling and QoS management