Upload
mapr-technologies
View
256
Download
1
Embed Size (px)
DESCRIPTION
These slides are from a talk Ted Dunning gave at Lawrence Livermore Labs in 2011. The talk gives an architectural outline of the MapR system and then discusses how this architecture facilitates large scale machine learning algorithms.
Citation preview
04/10/2023 © MapR Confidential 1
MapR Architecture and Machine Learning
1
04/10/2023 © MapR Confidential 2
Outline
• MapR system overview• Map-reduce review• MapR architecture• Performance Results• Map-reduce on MapR
• Machine learning on MapR
04/10/2023 © MapR Confidential 3
Map-Reduce
Input Output
Shuffle
04/10/2023 © MapR Confidential 4
Bottlenecks and Issues
• Read-only files• Many copies in I/O path• Shuffle based on HTTP• Can’t use new technologies• Eats file descriptors
• Spills go to local file space• Bad for skewed distribution of sizes
04/10/2023 © MapR Confidential 5
MapR Improvements
• Faster file system• Fewer copies• Multiple NICS• No file descriptor or page-buf competition
• Faster map-reduce• Uses distributed file system• Direct RPC to receiver• Very wide merges
04/10/2023 © MapR Confidential 6
MapR Innovations
• Volumes• Distributed management• Data placement
• Read/write random access file system• Allows distributed meta-data• Improved scaling• Enables NFS access
• Application-level NIC bonding• Transactionally correct snapshots and mirrors
04/10/2023 © MapR Confidential 7
Each container contains Directories & files Data blocks
Replicated on servers No need to manage
directly
MapR's ContainersFiles/directories are sharded into blocks, whichare placed into mini NNs (containers ) on disks
Containers are 16-32 GB segments of disk, placed on nodes
04/10/2023 © MapR Confidential 8
Container locations and replication
CLDB
N1, N2
N3, N2
N1, N2
N1, N3
N3, N2
N1
N2
N3
Container location database (CLDB) keeps track of nodes hosting each container
04/10/2023 © MapR Confidential 9
Containers represent 16 - 32GB of data Each can hold up to 1 Billion files and directories 100M containers = ~ 2 Exabytes (a very large cluster)
250 bytes DRAM to cache a container 25GB to cache all containers for 2EB cluster
But not necessary, can page to disk Typical large 10PB cluster needs 2GB
Container-reports are 100x - 1000x < HDFS block-reports Serve 100x more data-nodes Increase container size to 64G to serve 4EB cluster
Map/reduce not affected
MapR Scaling
04/10/2023 © MapR Confidential 10
MapR's Streaming Performance
Read Write0
250
500
750
1000
1250
1500
1750
2000
2250
Read Write0
250
500
750
1000
1250
1500
1750
2000
2250
HardwareMapRHadoopMB
persec
Tests: i. 16 streams x 120GB ii. 2000 streams x 1GB
11 x 7200rpm SATA 11 x 15Krpm SAS
Higher is better
04/10/2023 © MapR Confidential 11
Terasort on MapR
1.0 TB0
10
20
30
40
50
60
3.5 TB0
50
100
150
200
250
300
MapRHadoop
Elapsed time (mins)
10+1 nodes: 8 core, 24GB DRAM, 11 x 1TB SATA 7200 rpm
Lower is better
04/10/2023 © MapR Confidential 12
MUCH faster for some operations
# of files (millions)
Teststopped
hereCreateRate
Same 10 nodes …
04/10/2023 © MapR Confidential 14
NFS mounting models
• Export to the world• NFS gateway runs on selected gateway hosts
• Local server• NFS gateway runs on local host• Enables local compression and check summing
• Export to self• NFS gateway runs on all data nodes, mounted
from localhost
04/10/2023 © MapR Confidential 15
Export to the world
NFSServerNFS
ServerNFSServerNFS
ServerNFSClient
04/10/2023 © MapR Confidential 16
Client
NFSServer
Local server
Application
Cluster Nodes
04/10/2023 © MapR Confidential 17
ClusterNode
NFSServer
Universal export to self
Application
Cluster Nodes
04/10/2023 © MapR Confidential 18
ClusterNode
NFSServer
Application
ClusterNode
NFSServer
Application
ClusterNode
NFSServer
Application
Nodes are identical
04/10/2023 © MapR Confidential 19
Sharded text indexing
• Mapper assigns document to shard• Shard is usually hash of document id
• Reducer indexes all documents for a shard• Indexes created on local disk• On success, copy index to DFS• On failure, delete local files
• Must avoid directory collisions • can’t use shard id!
• Must manage local disk space
04/10/2023 © MapR Confidential 20
Conventional data flows
MapReducer
Input documents
Localdisk Search
EngineLocal
disk
Clustered index storage
Failure of a reducer causes garbage to accumulate in the
local disk
Failure of search engine requires
another download of the index from clustered storage.
04/10/2023 © MapR Confidential 21
SearchEngine
Simplified NFS data flows
MapReducer
Input documents
Clustered index storage
Failure of a reducer is cleaned up by
map-reduce framework
Search engine reads mirrored index directly.
04/10/2023 © MapR Confidential 22
Application to machine learning
• So now we have the hammer
• Let’s see some nails!
04/10/2023 © MapR Confidential 23
K-means
• Classic E-M based algorithm• Given cluster centroids,• Assign each data point to nearest centroid• Accumulate new centroids• Rinse, lather, repeat
04/10/2023 © MapR Confidential 24
Aggregatenew
centroids
K-means, the movie
Assignto
Nearestcentroid
Centroids
Input
04/10/2023 © MapR Confidential 25
But …
04/10/2023 © MapR Confidential 26
Averagemodels
Parallel Stochastic Gradient Descent
Trainsub
model
Model
Input
04/10/2023 © MapR Confidential 27
Updatemodel
Variational Dirichlet Assignment
Gathersufficientstatistics
Model
Input
04/10/2023 © MapR Confidential 28
Old tricks, new dogs
• Mapper• Assign point to cluster• Emit cluster id, (1, point)
• Combiner and reducer• Sum counts, weighted sum of points• Emit cluster id, (n, sum/n)
• Output to HDFS
Read fromHDFS to local disk by distributed cache
Written by map-reduce
Read from local disk from distributed cache
04/10/2023 © MapR Confidential 29
Old tricks, new dogs
• Mapper• Assign point to cluster• Emit cluster id, 1, point
• Combiner and reducer• Sum counts, weighted sum of points• Emit cluster id, n, sum/n
• Output to HDFSMapR FS
Read fromNFS
Written by map-reduce
04/10/2023 © MapR Confidential 30
Click modeling architecture
Featureextraction
anddown
sampling
Input
Side-data
Datajoin
SequentialSGD
Learning
Map-reduce
Now via NFS
04/10/2023 © MapR Confidential 31
Poor man’s Pregel
• Mapper
• Lines in bold can use conventional I/O via NFS
31
while not done: read and accumulate input models for each input: accumulate model write model synchronize reset input formatemit summary
04/10/2023 © MapR Confidential 32
Trivial visualization interface
• Map-reduce output is visible via NFS
• Legacy visualization just works
$ R> x <- read.csv(“/mapr/my.cluster/home/ted/data/foo.out”)> plot(error ~ t, x)> q(save=‘n’)
04/10/2023 © MapR Confidential 33
Conclusions
• We used to know all this• Tab completion used to work• 5 years of work-arounds have clouded our
memories
• We just have to remember the future