Huynh Tu Dang, and Fabien Hermenier. 9th Workshop on Hot topics in Dependable Systems (HotDep 2013). Farmington, USA. ACM. Full paper http://dl.acm.org/citation.cfm?id=2524226

1. Higher SLA Satisfaction in Datacenters
with Continuous Placement Constraints
Huynh Tu Dang
hdang@polytech.unice.fr
Fabien Hermenier
fabien.hermenier@unice.fr

2. SLA for a virtualised application

3. SLA for a virtualised application
spread the replicas

4. SLA for a virtualised application
performance guarantee

5. SLA for a virtualised application
low latency

6. reconfiguration algorithm
N2
N3
time
SLA: spread(VM1, VM2)
VM2
VM1
N1

7. reconfiguration algorithm
N2
N3
time
SLA: spread(VM1, VM2)
sys-admin query: offline(N1)
VM2
VM2
VM1
VM1
N1
Reconfiguration
process

8. reconfiguration algorithm
N2
N3
time
SLA: spread(VM1, VM2)
sys-admin query: offline(N1)
VM2
VM2
VM1
VM1
N1
with discrete restrictions

9. Discrete restriction is not enough
not an unpredictable situation,
an algorithmic issue

10. Evaluating the reliability of discrete
placement constraints
• simulate a 256-server datacenter
• running 350 HA webapp (5,200 VMs)
• BtrPlace as the reconfiguration algorithm
• 4 reconfiguration scenarios that mimic industrial use case
• 100 instances per scenario

11. Studied constraints
spread
among
splitAmong
maxOnline
singleResource
Capacity
replicas on distinct servers for fault
tolerance
DBs on a same edge-switch for a
fast synchronisation.
webapp split over 2 clusters
for disaster recovery
240 nodes online at maximum
to fit licensing policy
keep resource for hypervisor
management operations

12. scenario
vertical elasticity
Tiers 1
Tiers 2
Tiers 3

13. scenario
vertical elasticity
Tiers 1
Tiers 2
Tiers 3

14. scenariohorizontal elasticity
Tiers 1
Tiers 2
Tiers 3

15. scenariohorizontal elasticity
Tiers 1
Tiers 2
Tiers 3

16. scenarioboot storm x 400

17. scenarioserver failure

18. Migrations lead to unanticipated
Scenario Violated
SLAs
placements
Actions
VM Boot Migrate Node Boot Node Shutdown
Vertical Elasticity 40.72 0% 99.99% 0.005% 0.005%
Horizontal Elasticity 0.19 99.82% 0.18% 2.82% 0%
Server Failure 29.56 61.29% 35.89% 2.82% 0%
Boot Storm 0.35 98.57% 1.43% 0% 0%

19. Migrationstend to violate relative
100
75
50
25
0
Vertical
Elasticity
placement constraints
performance loss
Horizontal
Elasticity
Server
Failure
Boot
Storm
Violations
spread among splitAmong maxOnline
spof
failure

20. Trading unreliable
discrete constraints …
spread(VM[1,2])
we addressed an
assignment problem

21. … for safe
continuous constraints
spread(VM[1,2])
we must address
a scheduling problem

22. Continuous placement constraints
with

23. from discrete to continuous
among|simpleAmong
N1
N2 N3
N4 N5
N6
stay on a same partition by the end of
the reconfiguration process

24. from discrete to continuous
among|simpleAmong
N1
N2 N3
N4 N5
N6
stay on a same partition by the end of
the reconfiguration process

25. from discrete to continuous
among|simpleAmong
N1
N2 N3
N4 N5
N6
stay on a same partition by the end of
the reconfiguration process

26. from discrete to continuous
among|simpleAmong
N1
N2 N3
N4 N5
Disallow movements between partitions
• basic knowledge of a reconfiguration process
• still an assignment problem
N6

27. from discrete to continuous
among|simpleAmong
N1
N2 N3
N4 N5
Disallow movements between partitions
• basic knowledge of a reconfiguration process
• still an assignment problem
N6

28. continuous spread
discrete spread(VM[1,2]) ::= continuous spread(VM[1,2]) ::=
allDi↵erent(dhost
1 , dhost
2 )
allDi↵erent(dhost
1 , dhost
2 ) ^
dhost
1 = chost
2 =) astart
1 # aend
2 ^
dhost
2 = chost
1 =) astart
2 # aend
1
Disallow temporary overlapping
• require to know this may happen
• scheduling 101

29. continuous maxOnline
discrete maxOnline(N[1..10], 7)::=
continuous maxOnline(N[1..10], 7)::=
X10
i=1
nqi
 7
8i 2 [1, 10], non
i =
⇢
0 if nq
i = 1
astart
i otherwise
noff
i =
⇢
max(T) if nq
i = 0
aend
i otherwise
i # t ^ noff
8t 2 T, card({i|non
i })  7
scheduling 201
detailed knowledge of a
reconfiguration process
harder to imagine,
model & implement

30. Performance overhead
5 10 20 50 100
Duration (sec.)
Solved instances
discrete
continuous
0 20 40 60 80 100
10 20 50 100
Duration (sec.)
Solved instances
discrete
continuous
0 20 40 60 80 100
0 20 40 60 80 100vertical elasticity
discrete
continuous
5 10 20 50 100 200
Duration (sec.)
5 10 20 50 100
Duration (sec.)
Solved instances
Solved instances
discrete
continuous
0 20 40 60 80 100 boot storm
horitzontal elasticity server failure

31. Performance overhead
5 10 20 50 100
Duration (sec.)
Solved instances
discrete
continuous
0 20 40 60 80 100
10 20 50 100
Duration (sec.)
Solved instances
discrete
continuous
0 20 40 60 80 100
0 20 40 60 80 100vertical elasticity
discrete
continuous
5 10 20 50 100 200
Duration (sec.)
5 10 20 50 100
Duration (sec.)
Solved instances
Solved instances
discrete
continuous
0 20 40 60 80 100 boot storm
horitzontal elasticity server failure

32. Conclusions
• discrete restriction is not enough
• continuous restriction is a solution
• a different view on the problem
• challenging, but still possible to implement

33. Future Work
• a broader range of constraints and objectives
• reducing performance overhead
• static analysis to detect un-necessary
continuous constraints
• controlled relaxation to handle hard situations

34. http://btrp.inria.fr
open source, 20+ placement constraints,
demo, tutorials, everything for reproducibility