Advanced big data processing frameworks have been proposed to harness the fast data transmission capability of remote direct memory access (RDMA) over InfiniBand and RoCE. However, with the introduction of the non-volatile memory (NVM), these designs along with the default execution models, like MapReduce and Directed Acyclic Graph (DAG), need to be re-assessed to discover the possibilities of further enhanced performance.
In this context, we propose an accelerated execution framework (NVMD) for MapReduce and DAG that leverages the benefits of NVM and RDMA. NVMD introduces novel features for MapReduce and DAG, such as a hybrid push and pull shuffle mechanism and dynamic adaptation to the network congestion. The design has been incorporated into Apache Hadoop and Tez. Performance results illustrate that NVMD can achieve up to 3.65x and 3.18x improvement for Hadoop and Tez, respectively. In this talk, we will also present NVM-aware HDFS design and its benefits for MapReduce, Spark, and HBase.
Speaker: Shashank Gugnani, PhD Student, Ohio State University
Big data processing meets non-volatile memory: opportunities and challenges
1. Big Data Processing meets Non-volatile Memory:
Opportunities and Challenges
Talk at DataWorks Summit San Jose 2018
by
Shashank Gugnani, Dhabaleswar K. (DK) Panda, Xiaoyi Lu
The Ohio State University
E-mail: gugnani.2@osu.edu, panda@cse.ohio-state.edu, luxi@cse.ohio-state.edu
2. DataWorks Summit San Jose‘18 2Network Based Computing Laboratory
• Substantial impact on designing data management and processing systems in multiple tiers
– Front-end data accessing and serving (Online)
• Memcached + DB (e.g. MySQL), HBase
– Back-end data analytics (Offline)
• HDFS, MapReduce, Spark
Big Data Management and Processing on Modern Clusters
3. DataWorks Summit San Jose‘18 3Network Based Computing Laboratory
Big Data Processing with Apache Big Data Analytics Stacks
• Major components included:
– MapReduce (Batch)
– Spark (Iterative and Interactive)
– HBase (Query)
– HDFS (Storage)
– RPC (Inter-process communication)
• Underlying Hadoop Distributed File
System (HDFS) used by MapReduce,
Spark, HBase, and many others
• Model scales but high amount of
communication and I/O can be further
optimized!
HDFS
MapReduce
Apache Big Data Analytics Stacks
User Applications
HBase
Hadoop Common (RPC)
Spark
4. DataWorks Summit San Jose‘18 4Network Based Computing Laboratory
Drivers of Modern HPC Cluster and Data Center Architecture
• Multi-core/many-core technologies
• Remote Direct Memory Access (RDMA)-enabled networking (InfiniBand and RoCE)
– Single Root I/O Virtualization (SR-IOV)
• Solid State Drives (SSDs), NVM, Parallel Filesystems, Object Storage Clusters
• Accelerators (NVIDIA GPGPUs and FPGAs)
High Performance Interconnects –
InfiniBand (with SR-IOV)
<1usec latency, 200Gbps Bandwidth>
Multi-/Many-core
Processors
Cloud CloudSDSC Comet TACC Stampede
Accelerators / Coprocessors
high compute density, high
performance/watt
>1 TFlop DP on a chip
SSD, NVMe-SSD, NVRAM
5. DataWorks Summit San Jose‘18 5Network Based Computing Laboratory
• RDMA for Apache Spark
• RDMA for Apache Hadoop 2.x
– Plugins available for Cloudera (CDH) and Hortonworks (HDP)
distributions
• RDMA for Apache HBase
• RDMA for Memcached
• http://hibd.cse.ohio-state.edu
• Users Base: 285 organizations from 34 countries
• More than 26,650 downloads from the project site
The High-Performance Big Data (HiBD) Project
Available for InfiniBand, RoCE,
and, Ethernet
Support for Singularity and Docker
Significant performance
improvement with ‘RDMA+DRAM’
What about RDMA + NVM?
7. DataWorks Summit San Jose‘18 7Network Based Computing Laboratory
Non-Volatile Memory (NVM) and NVMe-SSD
3D XPoint from Intel & Micron Samsung NVMe SSD Performance of PMC Flashtec NVRAM [*]
• Non-Volatile Memory (NVM) provides byte-addressability with persistence
• The huge explosion of data in diverse fields require fast analysis and storage
• NVMs provide the opportunity to build high-throughput storage systems for data-intensive
applications
• Storage technology is moving rapidly towards NVM
[*] http://www.enterprisetech.com/2014/08/06/ flashtec-nvram-15-million-iops-sub-microsecond- latency/
8. DataWorks Summit San Jose‘18 8Network Based Computing Laboratory
• Popular methods employed by recent works to emulate NVRAM performance
model over DRAM
• Two ways:
– Emulate byte-addressable NVRAM over DRAM
– Emulate block-based NVM device over DRAM
NVRAM Emulation based on DRAM
Application
Virtual File System
Block Device PCMDisk
(RAM-Disk + Delay)
DRAM
mmap/memcpy/msync
Application
Persistent Memory Library
Clflush + Delay
DRAM
pmem_memcpy_persist
(DAX)
Load/store
Load/Store
open/read/write/close
9. DataWorks Summit San Jose‘18 9Network Based Computing Laboratory
• NRCIO: NVM-aware RDMA-based Communication
and I/O Schemes
• NRCIO for Big Data Analytics
• NVMe-SSD based Big Data Analytics
• Conclusion and Q&A
Presentation Outline
10. DataWorks Summit San Jose‘18 10Network Based Computing Laboratory
Design Scope (NVRAM)
• Use as block device or memory?
• Software overhead with I/O system calls
• Usage of memory requires application re-design
• Slower than DRAM
• Write performance ~10x worse
• Read performance ~1.5x worse
• Cost
• Cost is more than DRAM
How can you modify existing systems to take advantage of
the non-volatility of NVRAM?
11. DataWorks Summit San Jose‘18 11Network Based Computing Laboratory
Design Scope (NVM for RDMA)
D-to-N over RDMA N-to-D over RDMA N-to-N over RDMA
D-to-N over RDMA: Communication buffers for client are allocated in DRAM; Server uses NVM
N-to-D over RDMA: Communication buffers for client are allocated in NVM; Server uses DRAM
N-to-N over RDMA: Communication buffers for client and server are allocated in NVM
DRAM NVM
HDFS-RDMA
(RDMADFSClient)
HDFS-RDMA
(RDMADFSServer)
Client
CPU
Server
CPU
PCIe
NIC
PCIe
NIC
Client Server
NVM DRAM
HDFS-RDMA
(RDMADFSClient)
HDFS-RDMA
(RDMADFSServer)
Client
CPU
Server
CPU
PCIePCIe
NIC NIC
Client Server
NVM NVM
HDFS-RDMA
(RDMADFSClient)
HDFS-RDMA
(RDMADFSServer)
Client
CPU
Server
CPU
PCIePCIe
NIC NIC
Client Server
D-to-D over RDMA: Communication buffers for client and server are allocated in DRAM (Common)
12. DataWorks Summit San Jose‘18 12Network Based Computing Laboratory
NVRAM-aware Communication in NRCIO
NRCIO Send/Recv over NVRAM NRCIO RDMA_Read over NVRAM
13. DataWorks Summit San Jose‘18 13Network Based Computing Laboratory
NRCIO Send/Recv Emulation over PCM
• Comparison of communication latency using NRCIO send/receive semantics over InfiniBand QDR
network and PCM memory
• High communication latencies due to slower writes to non-volatile persistent memory
• NVRAM-to-Remote-NVRAM (NVRAM-NVRAM) => ~10x overhead vs. DRAM-DRAM
• DRAM-to-Remote-NVRAM (DRAM-NVRAM) => ~8x overhead vs. DRAM-DRAM
• NVRAM-to-Remote-DRAM (NVRAM-DRAM) => ~4x overhead vs. DRAM-DRAM
0
20000
40000
60000
80000
100000
120000
140000
1 2 4 8 16 32 64 128 256 512 1K 2K 4K 8K 16K 32K 64K 128K 256K 512K 1M 2M
Latency(us)
IB Send/Recv Emulation over PCM
DRAM-DRAM NVRAM-DRAM
DRAM-NVRAM NVRAM-NVRAM
DRAM-SSD-Msync_Per_Line DRAM-SSD-Msync_Per_8KPage
DRAM-SSD-Msync_Per_4KPage
14. DataWorks Summit San Jose‘18 14Network Based Computing Laboratory
NRCIO RDMA-Read Emulation over PCM
0
20000
40000
60000
80000
100000
120000
140000
1 2 4 8 16 32 64 128 256 512 1K 2K 4K 8K 16K 32K 64K 128K 256K 512K 1M 2M
Latency(us)
Message Size (bytes)
IB RDMA_READ Emulation over PCM
DRAM-DRAM NVRAM-DRAM
DRAM-NVRAM NVRAM-NVRAM
DRAM-SSD-Msync_Per_Line DRAM-SSD-Msync_Per_8KPage
DRAM-SSD-Msync_Per_4KPage
• Communication latency with NRCIO RDMA-Read over InfiniBand QDR + PCM memory
• Communication overheads for large messages due to slower writes into NVRAM from
remote memory; similar to Send/Receive
• RDMA-Read outperforms Send/Receive for large messages; as observed for DRAM-DRAM
15. DataWorks Summit San Jose‘18 15Network Based Computing Laboratory
• NRCIO: NVM-aware RDMA-based Communication
and I/O Schemes
• NRCIO for Big Data Analytics
• NVMe-SSD based Big Data Analytics
• Conclusion and Q&A
Presentation Outline
16. DataWorks Summit San Jose‘18 16Network Based Computing Laboratory
• Files are divided into fixed sized blocks
– Blocks divided into packets
• NameNode: stores the file system namespace
• DataNode: stores data blocks in local storage
devices
• Uses block replication for fault tolerance
– Replication enhances data-locality and read
throughput
• Communication and I/O intensive
• Java Sockets based communication
• Data needs to be persistent, typically on
SSD/HDD
NameNode
DataNodes
Client
Opportunities of Using NVRAM+RDMA in HDFS
17. DataWorks Summit San Jose‘18 17Network Based Computing Laboratory
Design Overview of NVM and RDMA-aware HDFS (NVFS)
• Design Features
• RDMA over NVM
• HDFS I/O with NVM
• Block Access
• Memory Access
• Hybrid design
• NVM with SSD as a hybrid
storage for HDFS I/O
• Co-Design with Spark and HBase
• Cost-effectiveness
• Use-case
Applications and Benchmarks
Hadoop MapReduce Spark HBase
Co-Design
(Cost-Effectiveness, Use-case)
RDMA
Receiver
RDMA
Sender
DFSClient
RDMA
Replicator
RDMA
Receiver
NVFS
-BlkIO
Writer/Reader
NVM
NVFS-
MemIO
SSD SSD SSD
NVM and RDMA-aware HDFS (NVFS)
DataNode
N. S. Islam, M. W. Rahman , X. Lu, and D. K.
Panda, High Performance Design for HDFS with
Byte-Addressability of NVM and RDMA, 24th
International Conference on Supercomputing
(ICS), June 2016
18. DataWorks Summit San Jose‘18 18Network Based Computing Laboratory
• RDMA communication buffers
allocated in NVRAM via D-to-N
approach
• buffers used in a round-robin
manner
• Data received in NVRAM buffers are
encapsulated in java I/O streams
• written to NVM through HDFS
file I/O APIs
• mmap files during HDFS I/O
Proposed Design (NVFS-BlkIO)
NVM
Communication Buffers
RDMADFSServer (DataNode)
Blk1
Blk1RDMADFSClient
Blk2RDMADFSClient
Blk3RDMADFSClient
BlkNRDMADFSClient
Blk2 Blk3
Blk4 … BlkN
write
read
write
read
write
read
write
read
.
.
.
Block Access
Memory Access
19. DataWorks Summit San Jose‘18 19Network Based Computing Laboratory
• Shared buffers for communication
and storage
• Buffers allocated in NVM; managed
in a linked list
• Data packets added to the tail
(first free buffer)
• Does not maintain HDFS block
structures
• Fast write, suboptimal read for
large number of blocks
• Needs all the buffers to be
registered in the communication
library
Proposed Design (NVFS-MemIO)
NVM
Free Buffers
RDMADFSServer (DataNode)
Blk1RDMADFSClient
Blk2RDMADFSClient
Blk3RDMADFSClient
BlkNRDMADFSClient
write
read
write
read
write
read
write
read
.
.
.
…
Used Buffers
Blk1 Pkt1 Data Nxt
Blk2 Pkt1 Data Nxt
Blk1 Pkt2 Data Nxt
BlkN Pkt1 Data Nxt
Flat Architecture
20. DataWorks Summit San Jose‘18 20Network Based Computing Laboratory
• Re-usable communication buffers
• Separate storage buffers in NVM
• Maintains HDFS block structures
• Overlapping between
communication and storage
– Efficient write
• Constant time (amortized) retrieval
of blocks
– Faster read over Flat Architecture
• Modularity
Proposed Design (NVFS-MemIO)
NVM
Communication Buffers
RDMADFSServer (DataNode)
Blk1RDMADFSClient
Blk2RDMADFSClient
Blk3RDMADFSClient
BlkNRDMADFSClient
write
read
write
read
write
read
write
read
.
.
.
…
Storage Buffers
Blk1
Pkt1 Data
BlkN
Pkt1 Data
PktM Data PktM Data
Blk1 Blk2 … BlkN
HashTable
LinkedList LinkedList
Block Architecture
21. DataWorks Summit San Jose‘18 21Network Based Computing Laboratory
Evaluation with Hadoop MapReduce
0
50
100
150
200
250
300
350
Write Read
AverageThroughput(MBps)
HDFS (56Gbps)
NVFS-BlkIO (56Gbps)
NVFS-MemIO (56Gbps)
• TestDFSIO on SDSC Comet (32 nodes)
– Write: NVFS-MemIO gains by 4x over
HDFS
– Read: NVFS-MemIO gains by 1.2x over
HDFS
TestDFSIO
0
200
400
600
800
1000
1200
1400
Write Read
AverageThroughput(MBps)
HDFS (56Gbps)
NVFS-BlkIO (56Gbps)
NVFS-MemIO (56Gbps)
4x
1.2x
4x
2x
SDSC Comet (32 nodes) OSU Nowlab (4 nodes)
• TestDFSIO on OSU Nowlab (4 nodes)
– Write: NVFS-MemIO gains by 4x over
HDFS
– Read: NVFS-MemIO gains by 2x over
HDFS
22. DataWorks Summit San Jose‘18 22Network Based Computing Laboratory
Evaluation with HBase
0
100
200
300
400
500
600
700
800
8:800K 16:1600K 32:3200K
Throughput(ops/s)
Cluster Size : No. of Records
HDFS (56Gbps) NVFS (56Gbps)
HBase 100% insert
0
200
400
600
800
1000
1200
8:800K 16:1600K 32:3200K
Throughput(ops/s)
Cluster Size : Number of Records
HBase 50% read, 50% update
• YCSB 100% Insert on SDSC Comet (32 nodes)
– NVFS-BlkIO gains by 21% by storing only WALs to NVM
• YCSB 50% Read, 50% Update on SDSC Comet (32 nodes)
– NVFS-BlkIO gains by 20% by storing only WALs to NVM
20%21%
23. DataWorks Summit San Jose‘18 23Network Based Computing Laboratory
Opportunities to Use NVRAM+RDMA in MapReduce
Disk Operations
• Map and Reduce Tasks carry out the total job execution
– Map tasks read from HDFS, operate on it, and write the intermediate data to local disk (persistent)
– Reduce tasks get these data by shuffle from NodeManagers, operate on it and write to HDFS (persistent)
• Communication and I/O intensive; Shuffle phase uses HTTP over Java Sockets; I/O operations take
place in SSD/HDD typically
Bulk Data Transfer
24. DataWorks Summit San Jose‘18 24Network Based Computing Laboratory
Opportunities to Use NVRAM in MapReduce-RDMA
DesignInputFiles
OutputFiles
IntermediateData
Map Task
Read Map
Spill
Merge
Map Task
Read Map
Spill
Merge
Reduce Task
Shuffle Reduce
In-
Mem
Merge
Reduce Task
Shuffle Reduce
In-
Mem
Merge
RDMA
All Operations are In-
Memory
Opportunities exist to
improve the
performance with
NVRAM
25. DataWorks Summit San Jose‘18 25Network Based Computing Laboratory
NVRAM-Assisted Map Spilling in MapReduce-RDMA
InputFiles
OutputFiles
IntermediateData
Map Task
Read Map
Spill
Merge
Map Task
Read Map
Spill
Merge
Reduce Task
Shuffle Reduce
In-
Mem
Merge
Reduce Task
Shuffle Reduce
In-
Mem
Merge
RDMA
NVRAM
Minimizes the disk operations in Spill phase
M. W. Rahman, N. S. Islam, X. Lu, and D. K. Panda, Can Non-Volatile Memory Benefit MapReduce Applications on HPC Clusters? PDSW-DISCS, with SC 2016.
M. W. Rahman, N. S. Islam, X. Lu, and D. K. Panda, NVMD: Non-Volatile Memory Assisted Design for Accelerating MapReduce and DAG Execution Frameworks on
HPC Systems? IEEE BigData 2017.
26. DataWorks Summit San Jose‘18 26Network Based Computing Laboratory
Comparison with Sort and TeraSort
• RMR-NVM achieves 2.37x benefit for Map
phase compared to RMR and MR-IPoIB;
overall benefit 55% compared to MR-IPoIB,
28% compared to RMR
2.37x
55%
2.48x
51%
• RMR-NVM achieves 2.48x benefit for Map
phase compared to RMR and MR-IPoIB;
overall benefit 51% compared to MR-IPoIB,
31% compared to RMR
27. DataWorks Summit San Jose‘18 27Network Based Computing Laboratory
Evaluation of Intel HiBench Workloads
• We evaluate different HiBench
workloads with Huge data sets
on 8 nodes
• Performance benefits for
Shuffle-intensive workloads
compared to MR-IPoIB:
– Sort: 42% (25 GB)
– TeraSort: 39% (32 GB)
– PageRank: 21% (5 million pages)
• Other workloads:
– WordCount: 18% (25 GB)
– KMeans: 11% (100 million samples)
28. DataWorks Summit San Jose‘18 28Network Based Computing Laboratory
Evaluation of PUMA Workloads
• We evaluate different PUMA
workloads on 8 nodes with
30GB data size
• Performance benefits for
Shuffle-intensive workloads
compared to MR-IPoIB :
– AdjList: 39%
– SelfJoin: 58%
– RankedInvIndex: 39%
• Other workloads:
– SeqCount: 32%
– InvIndex: 18%
29. DataWorks Summit San Jose‘18 29Network Based Computing Laboratory
• NRCIO: NVM-aware RDMA-based Communication
and I/O Schemes
• NRCIO for Big Data Analytics
• NVMe-SSD based Big Data Analytics
• Conclusion and Q&A
Presentation Outline
30. DataWorks Summit San Jose‘18 30Network Based Computing Laboratory
Overview of NVMe Standard
• NVMe is the standardized interface
for PCIe SSDs
• Built on ‘RDMA’ principles
– Submission and completion I/O
queues
– Similar semantics as RDMA send/recv
queues
– Asynchronous command processing
• Up to 64K I/O queues, with up to 64K
commands per queue
• Efficient small random I/O operation
• MSI/MSI-X and interrupt aggregation
NVMe Command Processing
Source: NVMExpress.org
31. DataWorks Summit San Jose‘18 31Network Based Computing Laboratory
Overview of NVMe-over-Fabric
• Remote access to flash with NVMe
over the network
• RDMA fabric is of most importance
– Low latency makes remote access
feasible
– 1 to 1 mapping of NVMe I/O queues
to RDMA send/recv queues
NVMf Architecture
I/O
Submission
Queue
I/O
Completion
Queue
RDMA Fabric
SQ RQ
NVMe
Low latency
overhead compared
to local I/O
32. DataWorks Summit San Jose‘18 32Network Based Computing Laboratory
Overview of SPDK
• High-performance storage library over NVMe
• Significant performance improvement over POSIX NVMe driver
33. DataWorks Summit San Jose‘18 33Network Based Computing Laboratory
Design Challenges with NVMe-SSD
• QoS
– Hardware-assisted QoS
• Persistence
– Flushing buffered data
• Performance
– Consider flash related design aspects
– Read/Write performance skew
– Garbage collection
• Virtualization
– SR-IOV hardware support
– Namespace isolation
• New software systems
– Disaggregated Storage with NVMf
– Persistent Caches
Co-design
34. DataWorks Summit San Jose‘18 34Network Based Computing Laboratory
Evaluation with RocksDB
0
5
10
15
Insert Overwrite Random Read
Latency (us)
POSIX SPDK
0
100
200
300
400
500
Write Sync Read Write
Latency (us)
POSIX SPDK
• 20%, 33%, 61% improvement for Insert, Write Sync, and Read Write
• Overwrite: Compaction and flushing in background
– Low potential for improvement
• Read: Performance much worse; Additional tuning/optimization required
35. DataWorks Summit San Jose‘18 35Network Based Computing Laboratory
Evaluation with RocksDB
0
5000
10000
15000
20000
Write Sync Read Write
Throughput (ops/sec)
POSIX SPDK
0
100000
200000
300000
400000
500000
600000
Insert Overwrite Random Read
Throughput (ops/sec)
POSIX SPDK
• 25%, 50%, 160% improvement for Insert, Write Sync, and Read Write
• Overwrite: Compaction and flushing in background
– Low potential for improvement
• Read: Performance much worse; Additional tuning/optimization required
36. DataWorks Summit San Jose‘18 36Network Based Computing Laboratory
QoS-aware SPDK Design
0
50
100
150
1 5 9 13 17 21 25 29 33 37 41 45 49
Bandwidth(MB/s)
Time
Scenario 1
High Priority Job (WRR) Medium Priority Job (WRR)
High Priority Job (OSU-Design) Medium Priority Job (OSU-Design)
0
1
2
3
4
5
2 3 4 5
JobBandwidthRatio
Scenario
Synthetic Application Scenarios
SPDK-WRR OSU-Design Desired
• Synthetic application scenarios with different QoS requirements
– Comparison using SPDK with Weighted Round Robbin NVMe arbitration
• Near desired job bandwidth ratios
• Stable and consistent bandwidth
S. Gugnani, X. Lu, and D. K. Panda, Analyzing, Modeling, and
Provisioning QoS for NVMe SSDs, (Under Review)
37. DataWorks Summit San Jose‘18 37Network Based Computing Laboratory
Conclusion and Future Work
• Exploring NVM-aware RDMA-based Communication and I/O Schemes
for Big Data Analytics
• Proposed a new library, NRCIO (work-in-progress)
• Re-design HDFS storage architecture with NVRAM
• Re-design RDMA-MapReduce with NVRAM
• Design Big Data analytics stacks with NVMe and NVMf protocols
• Results are promising
• Further optimizations in NRCIO
• Co-design with more Big Data analytics frameworks
38. DataWorks Summit San Jose‘18 38Network Based Computing Laboratory
{panda, luxi}@cse.ohio-state.edu
http://www.cse.ohio-state.edu/~panda
http://www.cse.ohio-state.edu/~luxi
Thank You!
Network-Based Computing Laboratory
http://nowlab.cse.ohio-state.edu/
The High-Performance Big Data Project
http://hibd.cse.ohio-state.edu/