HAL Id: hal-01875999
https://hal.inria.fr/hal-01875999
Submitted on 18 Sep 2018
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Latency-Aware Strategies for Placing Data Stream Analytics onto Edge Computing
Alexandre da Silva Veith, Marcos Dias de Assuncao, Laurent Lefèvre
To cite this version:
Alexandre da Silva Veith, Marcos Dias de Assuncao, Laurent Lefèvre. Latency-Aware Strategies for Placing Data Stream Analytics onto Edge Computing. HotEdge’18 - USENIX Workshop on Hot Topics in Edge Computing, Jul 2018, Boston, United States. pp.1, 2018. �hal-01875999�
Acknowledgment: This work was performed within the framework of the LABEX MILYON (ANR-10-LABX-0070) of Université de Lyon, within the program "Investissements d'Avenir" (ANR-11-IDEX-0007)
operated by the French National Research Agency (ANR).
Latency-Aware Strategies for Placing Data Stream Analytics
onto Edge Computing
Alexandre da Silva Veith, Marcos Dias de Assunção, Laurent Lefèvre
Avalon, LIP, ENS Lyon, University of Lyon - Lyon, France [email protected]
Edge device: Two cloudlets with 20
Raspberry PI 2 (4,74 MIPS at 1GHz and 1GB of RAM)
Cloud: Two AMD RYZEN 7 1800x
(304,51 MIPS at 3.6GHz and 1TB of RAM)
Microbenchmarks - Evaluate the impact of fork/join operators
Complex Applications - Investigate the outcomes for generic and multiple path applications
Experimental setup Analysis
Our approach: Microbenchmarks - 95% better Complex Applications - 50% faster in average - 52% of reduction in the communication for sinks located in the edge
Reference
[1] Taneja, M., Davy, A.:
Resource aware placement of IoT application modules in fog-cloud computing paradigm. 2017
Cloudlet 1
Internet
Cloudlet 2
Edge
devices
Gateway
Cloud
LAN
WAN
Scenarios Monitoring of operational infrastructureAnomaly detection, fraud detection
Social networks
Internet of Things (IoT)
Applications
Directed dataflows whose vertices are operators that execute a function over the incoming data and edges define how data flows between them
Operator Characteristics
Path 1
Path 2
Selectivity:
The ratio of number of
input messages to output messages
n messages
m messages
Data compression/ expansion factor:
The ratio of the size of input events
to the size of output events
n bytes
m bytes
Fork/ Split: messages can be replicated
or scheduled to downstream operators
Merge/ Join: merges the outcome of
upstream operators
Paths and location:
multiple sources and sinks distributed across cloud and/or edge Latency-Aware Decisions Minimize the response times in all paths Application Splitting Across Edge and Cloud
Operator Characteristics and Requirements (CPU, Memory, Bandwidth)
Edge and Cloud Capabilities (CPU, Memory,
Bandwidth, Latency)
Edge Cloud
Response time (RT): the total time taken for a
message to traverse a path
Resource and Application Models
Resource 1
Op. 1
Message
Queue
Data transfer
service
Op. 2
Resource 2
Op. 3
Op. 4
Messages
Network
Operators - communication and computation services Services - M/M/1 queues
Not directly connected
Directly connected
Example of operator placement
LAN:
- Latency: U(0.015-0.8)ms - Bandwidth: 100 Mbps
Conclusions
- Dynamically movement of operators across edge
and cloud
WAN:
- Latency: U(65-85)ms - Bandwidth: 1 Gbps
- 50% faster in generic and complex dataflows
Finding Application Patterns
11
2 10
1
7
6
4
9
3
0
5
8
Initial setup Phase 1 Pattern recog. Phase 2Transform hierarchySeries
Phase 3 Phase 4
2
11
10
1
7
6
4
9
3
0
5
8
10
1
11
2
7
6
9
4
10 5
3
5
11
10
0
11 8
0
5
4
0
Response Time Rate (RTR) Strategy
- Find the dataflow split points (green)
- Edge (red) and cloud (blue) placements
2) For each operator in the sequence
- Response time estimation for all resources - Host with shortest response time is elected to host the operator
1) BFS-Traversal algorithm = operator sequence
7
6
4
9
3
5
11
2
11
0
8
10
5
0
5
10
1
11
10
4
0
RTR with Region Patterns (RTR+RP) Strategy Final
placement
Operator sequence = RTR
Candidate placement = destination infrastructure of the message
- Only cloud = cloud
- Edge and/or cloud = edge
Response time estimation - only for edge candidates Edge Cloud Edge does not meet the constraints Regions definition
