+ All Categories
Home > Documents > Accelerating Anomaly Detection Algorithms on FPGA-Based...

Accelerating Anomaly Detection Algorithms on FPGA-Based...

Date post: 20-May-2020
Category:
Upload: others
View: 25 times
Download: 0 times
Share this document with a friend
4
33 RD I NTERNATIONAL COSMIC RAY CONFERENCE,RIO DE JANEIRO 2013 THE ASTROPARTICLE PHYSICS CONFERENCE Space Mission Lomonosov on Study Gamma-Ray Bursts and UHECR A.M. AMELUSHKIN 1 , V.V. BENGHIN 1 , V.V. BOGOMOLOV 1 , G.K. GARIPOV 1 , E.S. GORBOVSKOY 2 , B. GROSSAN 3 , A.F. I YUDIN 1 , B.A. KHRENOV 1 , P.A. KLIMOV 1 , J. LEE 4 , V.M. LIPUNOV 2 , G. NA 4 , V.I. OSEDLO 1 , M.I. PANASYUK 1 , I.H. PARK 4 , V.L. PETROV 1 , S.A. SHARAKIN 1 ,YU.SHPRITS 5 , G.F. SMOOT 3 , S.I. SVERTILOV 1 , N.N. VEDENKIN 1 , I.V. YASHIN 1 1 Lomonosov Moscow State University, Skobeltsyn Institute of Nuclear Physics, Moscow, Russia 2 Lomonosov Moscow State University, Shternberg Astronomical Institute, Moscow, Russia 3 Berkeley Center for Cosmological Physics, Berkeley, California, USA 4 Department of Physics, Sungkyunkwan University, Seobu-ro, Jangangu, Suwonsi, Gyeongido, 440-746, Korea 5 Institute of Geophysics and Planetary Physics, UCLA, 405 Hilgard Ave / 7127, Los Angeles, CA, USA. [email protected] Abstract: The main idea of Lomonosov space mission is to study extreme astrophysical phenomena in the Universe, such as cosmic gamma-ray bursts (GRB) and ultra-high energy cosmic rays (UHECR). GRB being one of the most powerful events in the Universe occur not only in gamma-range, but also in optics and UV. Due to unusually powerful brightness of GRBs, studying of their properties allows the researchers to look in the epoch of early Universe, i.e. to study evolution of stars and stellar populations with red shift starting from z 0.1. Other extreme phenomenon in the Universe is a flux of UHECR, which is most likely produced in Active Galactic Nuclei (AGN). The fundamental problem is to estimate maximal particle energy, to which they could be accelerated in such sources. AGN are very distant objects, UHECR go a long way before coming to the Earth. During their propagation UHECR lose energy due to photo-production of secondary particles (mostly pions) on the microwave background photons. It leads to a natural limit of observable cosmic ray particle energy, Greisen- Zatsepin-Kuzmin limit,and to UHECR energy spectrum cut-off at energy of about 5 · 10 19 eV. Studies of mentioned above problems of extreme phenomena dictate scientific objectives of large scale space experiment Lomonosov with a specific set of instruments: detectors of GRB in wide range of wavelengths (in optics, ultra-violet, X-rays and gamma-rays) and large aperture telescope for recording fluorescence light from the atmosphere generated by UHECR. Main parameters and brief description of these instruments are presented. Keywords: Lomonosov, UHECR, GRB. 1 Introduction Studies of extremely high energy and power processes such as GRB and UHECR are of great importance not only for understanding these phenomena, but also for developing theory of the early Universe. GRBs are observed as short (from dozens of milliseconds up to dozens of seconds) increases of gamma-quanta flux with quanta energies from 10 5 eV up to at least 10 9 eV. Discovered in 60s years of 20th century they are still at the cutting edge of astrophysics. These phenomena being the most powerful in the Universe occur not only in gamma- range, but also in optics and UV. The power of the explosion of these most bright astrophysical objects achieves 10 51 10 53 erg/s. GRB optical emission lasts up to several hours or even days, it can be an evidence of afterglow which appears after a giant explosion in the external shock wave expanding in the interstellar space and stellar wind of the exploded star. Probably, it is a process of collapse of a fast-rotating very massive star to a black hole in the case of so-called long- duration (more than a few seconds) bursts or merging of a neutron stars in tight binary system in the case of so-called short-duration (less than a second) bursts. However, those models are under discussion and nature of this extraordinary phenomenon is still unknown. Due unusually powerful brightness of GRBs, studying of their properties allows researchers to look in epoch of the early Universe, i.e. to study evolution of stars and stellar populations within the wide range of red shift from z 0.1 up to z 15–20, which is more than 98% of age of the Universe. The other extreme phenomena in the Universe are ultra- high energy cosmic rays, which are most likely produced by the Active Galactic Nuclei (AGN). The fundamental prob- lem is to estimate maximal particle energy, to which they could be accelerated in such sources, and whether there is a maximum energy to which particles can be accelerated anywhere in the Universe. Because AGN are very distant objects, UHECR go a long way before coming to the Earth. During their propagation UHECR lose energy due to photo- production of secondary particles (mostly pions) on the mi- crowave background photons. It leads to a natural limit of observable cosmic ray particle energy and to UHECR en- ergy spectrum cut-off at the photo-production energy thresh- old, i.e. about 5 · 10 19 eV (Greisen-Zatsepin-Kuzmin cut- off). However, at present we have only limited and contra- dictory information from ground-based experimental arrays about the energy spectrum and composition of cosmic parti- cles at extremely high energies. UHECR detectors on board of satellites promise to get new and rich information in this interesting field of science [1, 2, 3, 4].
Transcript
Page 1: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

Accelerating Anomaly Detection Algorithms on

FPGA-Based High-Speed NICsHiroki Matsutani

Dept. of ICS, Keio Universityhttp://www.arc.ics.keio.ac.jp/~matutani

August 2nd, 2018 International Forum on MPSoC for Software-Defined Hardware (MPSoC'18) 1

Page 2: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

Accelerator design for big data

2

Input stream data Message queuing

RealtimeView

Database layer(Polyglot persistence)

DB queries

Data exchange (serialization)

Customer analysis Topic prediction Blockchain records Geolocation query …

Big data (Surveillance, Network service, SNS,UAV, IoT)

BatchView

BatchLearning

Inference

Batchprocessing

Streamprocessing

Page 3: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

Today’s talk: Online learning

3

Input stream data Message queuing

RealtimeView

Database layer(Polyglot persistence)

DB queries

Data exchange (serialization)

Customer analysis Topic prediction Blockchain records Geolocation query …

Big data (Surveillance, Network service, SNS,UAV, IoT)

BatchView

BatchLearning

I/O intensive Compute intensive

Message queuing middleware

(Apache Kafka)

Stream processing (Apache Spark

Streaming)

Batch processing(Apache Spark)

Batch learning (Apache Spark

MLlib)Serialization

(Apache Thrift)KVS / Column DB(Redis, HBase)

Document DB(MongoDB)

Graph DB (Neo4j), graph

processing

Inference(DNN, CNN, RNN)

Tight integration of I/O and compute FPGA

Massive parallelism Networked GPU cluster

GPUsHostSwitchFour

10GbE FPGA

Inference,Online

learning

Batchprocessing

Streamprocessing

Online learning(SGD, ChangeFinder,

OS-ELM)

Page 4: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

Offline vs. Online learning

4

All training data

New training

data+

Predictor

OK or NG?Test data

GPU-based batch processing

Examples: DNN, CNN, … Learning cost is high Predictor updated infrequently

Learning cost is low Predictor updated frequently Not very versatile

Predictor Test data ≒Training data

Sequential learning+ Inference

FPGA NIC/Switch

10GbEx4

FPGA-based stream processing

OK or NG?

Offline learning Online learning

Page 5: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

Online learning approaches

5

Next value Xt is predictedbased on recent p values

Online sequential learning for SLFN (input, hidden, and output layers)

Time t

Xt-1

Xt-2

Xt-3Xt-4

Xt-5

?

Xt= β1Xt-1 + β2Xt-2 + …

ChangeFinder:Outlier and change point detections on time-series data

AR-model based Neural network

Page 6: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

ChangeFinder on 10GbE FPGA• ChangeFinder algorithm

6

[J.Takeuchi, IEEE TKDE'06]

Elapsed time

Input data Xt

Change-point score

Step 1 (Outlier score):Receive input data Xt at time tCalculate outlier score of Xt based on past dataInfluence of past data controlled by discount rate r

Step 2 (Smoothing):Calculate moving average Yt of the outlier scoreSmoothing is controlled by window size S

Step 3 (Change-point score):Step 1 is performed for YtThe result is change-point score

Page 7: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

ChangeFinder on 10GbE FPGA• ChangeFinder algorithm

7

[J.Takeuchi, IEEE TKDE'06]

Elapsed time

Input data Xt

Change-point score

Threshold

Step 1 (Outlier score):Receive input data Xt at time tCalculate outlier score of Xt based on past dataInfluence of past data controlled by discount rate r

Step 2 (Smoothing):Calculate moving average Yt of the outlier scoreSmoothing is controlled by window size S

Step 3 (Change-point score):Step 1 is performed for YtThe result is change-point score

Page 8: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

ChangeFinder on 10GbE FPGA• 10GbE NIC datapath by Verilog HDL• Application logic in wrapper in HLS

8

10GMACRX 0

DMARX

Input Arbiter

Output PortLookup

BRAM OutputQueues

10GMACRX 3

10GMACTX 0

10GMACTX 3

DMATX

Wrapper

ChangeFinder.c

AXI4-Stream

AXI4-Stream

[T.Iwata, HeteroPar‘18]

Page 9: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

ChangeFinder on 10GbE FPGA• Throughput: 83.4% of 10GbE line rate

9

Youtube Video:https://www.youtube.com/watch?v=wgTcBfkE5hY

Page 10: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

Online learning approaches

10

Time t

Xt-1

Xt-2

Xt-3Xt-4

Xt-5

?

Xt= β1Xt-1 + β2Xt-2 + …

Next value Xt is predictedbased on recent p values

Online sequential learning for SLFN (input, hidden, and output layers)

n N m

Weight vector β00~β(N-1)(m-1)

ChangeFinder:Outlier and change point detections on time-series data

OS-ELM:Single hidden layer neural network (SLFN)

AR-model based Neural network

[N. Liang, TNN 2006]

Page 11: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

Online learning + unsupervised

Pre-trained predictor

OK or NG?Test data Learning

predictorTest data ≒Training data

Sequential learning+ Inference

Offline learning Online learning

Inference only

Normal values (incl. noise) are learned after the deployment Anomaly detection adapted to a given environment

[M.Tsukada, HeteroPar’18]*Collaboration with Prof. M.Kondo (UTokyo)

+Unsupervised anomaly detection

(No training data needed)

11

Page 12: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

Online learning + unsupervised• Learn vibration pattern of fan + noise

12

Youtube Video:https://www.youtube.com/watch?v=tCw7p7bjwTs

Page 13: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

Summary: Online learning FPGA

14

Input stream data Message queuing

RealtimeView

Database layer(Polyglot persistence)

DB queries

Data exchange (serialization)

Customer analysis Topic prediction Blockchain records Geolocation query …

Big data (Surveillance, Network service, SNS,UAV, IoT)

BatchView

BatchLearning

I/O intensive Compute intensive

Message queuing middleware

(Apache Kafka)

Stream processing (Apache Spark

Streaming)

Batch processing(Apache Spark)

Batch learning (Apache Spark

MLlib)Serialization

(Apache Thrift)KVS / Column DB(Redis, HBase)

Document DB(MongoDB)

Graph DB (Neo4j), graph

processing

Inference(DNN, CNN, RNN)

Tight integration of I/O and compute FPGA

Massive parallelism Networked GPU cluster

GPUsHostSwitchFour

10GbE FPGA

Inference,Online

learning

Batchprocessing

Stream processing

Online learning(SGD, ChangeFinder,

OS-ELM)

Page 14: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

References (1/2)• Outlier detection on 10GbE FPGA NIC

– Ami Hayashi, et.al., "An FPGA-Based In-NIC Cache Approach for Lazy Learning Outlier Filtering", PDP 2017.

– Ami Hayashi, et.al., "A Line Rate Outlier Filtering FPGA NIC using 10GbE Interface", ACM Comp Arch News (2015).

• Change-point detection on FPGA NIC– Takuma Iwata, et.al., "Accelerating Online

Change-Point Detection Algorithm using 10GbE FPGA NIC", HeteroPar 2018.

15

Page 15: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

References (2/2)• Online sequential unsupervised anomaly

detector on FPGA– Mineto Tsukada, et.al., "OS-ELM-FPGA: An

FPGA-Based Online Sequential Unsupervised Anomaly Detector", HeteroPar 2018.

16

Page 16: Accelerating Anomaly Detection Algorithms on FPGA-Based ...matutani/papers/matsutani_mpsoc2018.pdf(Apache Kafka) Stream processing (Apache Spark Streaming) Batch processing (Apache

17

Thank you !

Acknowledgement:This work is supported by JST CREST JPMJCR1785

Thank you for listening


Recommended