Autonomic SLA-driven Provisioning for Cloud Applications Nicolas Bonvin, Thanasis Papaioannou, Karl Aberer Presented by Ismail Alan

 This paper discusses an economic approach to managing cloud resources for individual applications based on established Service Level Agreements (SLA).  The approach attempts to mitigate the impact (to individual applications) of varying loads and random failures within the cloud. About the Paper

 A distributed, component-based application running on an elastic infrastructure Cloud Apps – Issue #1 : Placement 2 C1 C2C3C4

 A distributed, component-based application running on an elastic infrastructure Cloud Apps – Issue #1 : Placement 3 C2 VM1 C3 VM2 C4 VM3 C1

 A distributed, component-based application running on an elastic infrastructure Performance of C1, C2 and C3 is probably less than C4 No info on other VMs colocated on same server ! Cloud Apps – Issue #1 : Placement 5 No control on placement C3 VM2 C4 VM3 Server 2 C1 C2 Server 1 VM1

 Load-balanced traffic to 4 identical components on 4 identical VMs Cloud Apps – Issue #2 : Unstability 6 C1 VM2 100 ms C1 VM3 100 ms C1 VM4 100 ms C1 VM1 100 ms

 Load-balanced trafic to 4 identical components on 4 identical VMs – VM performance can vary because of different factors  Physical server, Hypervisor, Storage,... Cloud Apps – Issue #2 : Unstability 7 C1 VM2 140 ms C1 VM3 100 ms C1 VM4 100 ms C1 VM1 100 ms

 Load-balanced trafic to 4 identical components on 4 identical VMs – VM performance can vary because of different factors  Physical server, Hypervisor, Storage,... Component overloaded Cloud Apps – Issue #2 : Unstability 8 C1 VM1 130 ms C1 VM2 140 ms C1 VM3 100 ms C1 VM4 100 ms

 Load-balanced trafic to 4 identical components on 4 identical VMs – VM performance can vary because of different factors  Physical server, Hypervisor, Storage,... Component overloaded Component bug, crash, deadlock,... Cloud Apps – Issue #2 : Unstability 9 C1 VM1 130 ms C1 VM2 140 ms C1 VM3 100 ms C1 VM4 infinity

 Load-balanced trafic to 4 identical components on 4 identical VMs – VM performance can vary because of different factors  Physical server, Hypervisor, Storage,... Component overloaded Component bug, crash, deadlock,... Failure of C1 on VM4 -> load should be rebalanced Cloud Apps – Issue #2 : Unstability 10 C1 VM1 140 ms C1 VM2 150 ms C1 VM3 130 ms C1 VM4 infinity

 Load-balanced trafic to 4 identical components on 4 identical VMs – VM performance can vary because of different factors  Physical server, Hypervisor, Storage,... Component overloaded Component bug, crash, deadlock,... Failure of C1 on VM4 -> load should be rebalanced Cloud Apps – Issue #2 : Unstability 11 C1 VM1 140 ms C1 VM2 150 ms C1 VM3 130 ms C1 VM4 infinity Application should react early !

 Build for failures –––– Do not trust the underlying infrastructure Do not trust your components either !  Components should adapt to the changing conditions –––––– Quickly Automatically e.g. by replacing a wonky VM by a new one Cloud Apps – Overview 12

Scarce: a framework to build scalable cloud applications

Architecture Overview 14 Agent Server GOSSIPING + BROADCAST Agent A BEBE  An agent on each server / VM working based on Economic approach –––– starts/stops/monitors the components Takes decisions on behalf of the components  An agent communicates with other agents –––– Routing table Status of the server (resources usage) Agent

An economic approach 15  Time is split into epochs At each epoch servers charge a virtual rent for hosting a component according to –––––– Current resource usage (I/O, CPU,...) of the server Technical factors (HW, connectivity,...) Non-technical factors (location)

An economic approach 16  Components –––––– Pay virtual rent at each epoch Gain virtual money by processing requests Take decisions based on balance ( = gain – rent )  Replicate, migrate, suicide, stay  Virtual rents are updated by gossiping (no centralized board)  Time is split into epochs At each epoch servers charge a virtual rent for hosting a component according to –––––– Current resource usage (I/O, CPU,...) of the server Technical factors (HW, connectivity,...) Non-technical factors (location)

Economic model (i) 17 Balance of the component Utility of component Usage % of the server resources by component c Migration threshold Rent paid by component

Economic model (ii) 18  Based on the negative balance a component may migrate or stop  Calculate the availability  If satisfactory, the component stops.  Otherwise, try to find a less expensive server.  Based on the positive balance a component may replicate  Verify that can afford replication  If it can afford replication for consecutive epochs, replicate  Otherwise, continue to run.

Economic model (iii) 19  Choosing a candidate server j during replication/migration of a component i  netbenefit maximization  2 optimization goals :  high-availability by geographical diversity of replicas  low latency by grouping related components  g j : weight related to the proximity of the server location to the geographical distribution of the client requests to the component S i is the set of server hosting a replica of component I Diversity function returns geographical distance among each server pair. 

SLA Performance Guarantees (i) 20  Each component has its own SLA constraints SLA derived directly from entry components  Resp. Time = Service Time + max (Resp. Time of Dependencies) C1 SLA ::500ms C2 C3 C4 C5

SLA Performance Guarantees (ii) 21  SLA propagation from parents to children Parent j sends its performance constraints (e.g. response time upper bound) to its dependencies D(j) : Child i computes its own performance constraints : : group of constraints sent by the replicas of the parent g

SLA Performance Guarantees (iii) 22  SLA propagation from parents to children

Automatic Provisioning 23  Usage of allocated resources is maximized : –––– autonomic migration / replication / suicide of components not enough to ensure end-to-end response time  Each individual component has to satisfy its own SLA –––– SLA easily met -> decrease resources (scale down) SLA not met -> increase resources (scale up, scale out)  Cloud resources managed by framework via cloud API

Adaptivity to slow servers 24  Each component keeps statistics about its children – e.g. 95 th perc. response time  A routing coefficient is computed for each child at each epoch – Send more requests to more performant children

Evaluation

Evaluation: Setup 26  An application composed by 5 different components, mostly CPU-intensive  8 8-cores servers (Intel Core i7 920, 2.67 GHz, 8GB, Linux trunk-amd64) The components interact with the cloud infrastructure through an API Comparison of Scarce model with static approach. C1 SLA ::500ms C2 C3 C4 C5

Adaptation to Varying Load (i) 27  5 rps to 60 rps at minute 8, step 5 rps/min Static setup : 2 servers with 4 cores Fig 9 : Throughput of the application during the varying load experiments Fig 6 Scarce : Resources used by the app. over time for varying request load.

Adaptation to Varying Load (ii) 28  5 rps to 60 rps at minute 8, step 5 rps/min Static setup : 2 servers with 4 cores Fig 7: Mean response times of the application (SLA : 500 ms) as perceived by remote clients under the adaptive approach (“Scarce”) and the static setup. Fig 8: 95th percentile response times of the application (SLA : 500 ms) as perceived by remote clients under Scarce and the static setup.

Adaptation to Slow Server 29  Max 2 cores/server, 25 rps At minute 4, a server gets slower (200 ms delay) Fig 13: Resources used by the application over time in case of a “wonky” server. Fig. 12. Mean and 95th percentile response times of the application (SLA : 500 ms) as perceived by remote clients in case of a “wonky” server.

Scalability 30  Add 5 rps per minute until 150 rps Max 6 cores/server Fig 14: Mean and 95th percentile response times of the application (SLA:500ms) as perceived by remote clients in the scalability experiment. Fig 16: Resources used by the application over time during the scalability experiment. Fig 15: Scarce : Throughput of the application during the scalability experiment.

Conclusion

32  Framework for building cloud applications Elasticity : add/remove resources High Availability : software, hardware, network failures Scalability : growing load, peaks, scaling down,... – Quick replication of busy components  Load Balancing : load has to be shared by all available servers –––––– Replication of busy components Migration of less busy components Reach equilibrium when load is stable  SLA performance guarantees – Automatic provisioning  No synchronization, fully decentralized

Thank you !