Building Efficient HPC Clouds with MCAPICH2 and RDMA-Hadoop over SR-IOV InfiniBand Clusters

Building Efficient HPC Clouds with MVAPICH2 and
RDMA-Hadoop over SR-IOV InfiniBand Clusters
Talk at OpenFabrics Alliance Workshop (OFAW ‘17)
by
Dhabaleswar K. (DK) Panda
The Ohio State University
E-mail: panda@cse.ohio-state.edu
http://www.cse.ohio-state.edu/~panda
Xiaoyi Lu
The Ohio State University
E-mail: luxi@cse.ohio-state.edu
http://www.cse.ohio-state.edu/~luxi

OFAW 2017 2Network Based Computing Laboratory
• Cloud Computing focuses on maximizing the effectiveness of the shared resources
• Virtualization is the key technology for resource sharing in the Cloud
• Widely adopted in industry computing environment
• IDC Forecasts Worldwide Public IT Cloud Services Spending to Reach Nearly $108 Billion
by 2017 (Courtesy: http://www.idc.com/getdoc.jsp?containerId=prUS24298013)
Cloud Computing and Virtualization
VirtualizationCloud Computing

Drivers of Modern HPC Cluster and Cloud Architecture
• Multi-core/many-core technologies, Accelerators
• Large memory nodes
• Solid State Drives (SSDs), NVM, Parallel Filesystems, Object Storage Clusters
• Remote Direct Memory Access (RDMA)-enabled networking (InfiniBand and RoCE)
• Single Root I/O Virtualization (SR-IOV)
High Performance Interconnects –
InfiniBand (with SR-IOV)
<1usec latency, 200Gbps Bandwidth>
Multi-/Many-core
Processors
SSDs, Object Storage
Clusters
Large memory nodes
(Upto 2 TB)
Cloud CloudSDSC Comet TACC Stampede

• Single Root I/O Virtualization (SR-IOV) is providing new opportunities to design
HPC cloud with very little low overhead
Single Root I/O Virtualization (SR-IOV)
• Allows a single physical device, or a
Physical Function (PF), to present itself as
multiple virtual devices, or Virtual
Functions (VFs)
• VFs are designed based on the existing
non-virtualized PFs, no need for driver
change
• Each VF can be dedicated to a single VM
through PCI pass-through
• Work with 10/40 GigE and InfiniBand

• High-Performance Computing (HPC) has adopted advanced interconnects and protocols
– InfiniBand
– 10/40/100 Gigabit Ethernet/iWARP
– RDMA over Converged Enhanced Ethernet (RoCE)
• Very Good Performance
– Low latency (few micro seconds)
– High Bandwidth (100 Gb/s with EDR InfiniBand)
– Low CPU overhead (5-10%)
• OpenFabrics software stack with IB, iWARP and RoCE interfaces are driving HPC systems
• How to Build HPC Clouds with SR-IOV and InfiniBand for delivering optimal performance?
Building HPC Cloud with SR-IOV and InfiniBand

• Virtualization Support with Virtual Machines and Containers
– KVM, Docker, Singularity, etc.
• Communication coordination among optimized communication channels on Clouds
– SR-IOV, IVShmem, IPC-Shm, CMA, etc.
• Locality-aware communication
• Scalability for million to billion processors
– Support for highly-efficient inter-node and intra-node communication (both two-sided and one-sided)
• Scalable Collective communication
– Offload; Non-blocking; Topology-aware
• Balancing intra-node and inter-node communication for next generation nodes (128-1024 cores)
– Multiple end-points per node
• NUMA-aware communication for nested virtualization
• Integrated Support for GPGPUs and Accelerators
• Fault-tolerance/resiliency
– Migration support with virtual machines
• QoS support for communication and I/O
• Support for Hybrid MPI+PGAS programming (MPI + OpenMP, MPI + UPC, MPI + OpenSHMEM, MPI+UPC++, CAF, …)
• Energy-Awareness
• Co-design with resource management and scheduling systems on Clouds
– OpenStack, Slurm, etc.
Broad Challenges in Designing Communication and I/O Middleware
for HPC on Clouds

• High-Performance designs for Big Data middleware
– RDMA-based designs to accelerate Big Data middleware on high-performance Interconnects
– NVM-aware communication and I/O schemes for Big Data
– SATA-/PCIe-/NVMe-SSD support
– Parallel Filesystem support
– Optimized overlapping among Computation, Communication, and I/O
– Threaded Models and Synchronization
• Fault-tolerance/resiliency
– Migration support with virtual machines
– Data replication
• Efficient data access and placement policies
• Efficient task scheduling
• Fast deployment and automatic configurations on Clouds
Additional Challenges in Designing Communication and I/O
Middleware for Big Data on Clouds

• MVAPICH2-Virt with SR-IOV and IVSHMEM
– Standalone, OpenStack
• SR-IOV-enabled VM Migration Support in MVAPICH2
• MVAPICH2 with Containers (Docker and Singularity)
• MVAPICH2 with Nested Virtualization (Container over VM)
• MVAPICH2-Virt on SLURM
– SLURM alone, SLURM + OpenStack
• Big Data Libraries on Cloud
– RDMA-Hadoop, OpenStack Swift
Approaches to Build HPC Clouds

MVAPICH2 Software Family
High-Performance Parallel Programming Libraries
MVAPICH2 Support for InfiniBand, Omni-Path, Ethernet/iWARP, and RoCE
MVAPICH2-X Advanced MPI features, OSU INAM, PGAS (OpenSHMEM, UPC, UPC++, and CAF), and
MPI+PGAS programming models with unified communication runtime
MVAPICH2-GDR Optimized MPI for clusters with NVIDIA GPUs
MVAPICH2-Virt High-performance and scalable MPI for hypervisor and container based HPC cloud
MVAPICH2-EA Energy aware and High-performance MPI
MVAPICH2-MIC Optimized MPI for clusters with Intel KNC
Microbenchmarks
OMB Microbenchmarks suite to evaluate MPI and PGAS (OpenSHMEM, UPC, and UPC++)
libraries for CPUs and GPUs
Tools
OSU INAM Network monitoring, profiling, and analysis for clusters with MPI and scheduler
integration
OEMT Utility to measure the energy consumption of MPI applications

HPC on Cloud Computing Systems: Challenges Addressed by OSU
So Far
HPC and Big Data Middleware
Networking Technologies
(InfiniBand, Omni-Path, 1/10/40/100
GigE and Intelligent NICs)
Storage Technologies
(HDD, SSD, NVRAM, and NVMe-SSD)
HPC (MPI, PGAS, MPI+PGAS, MPI+OpenMP, etc.)
Applications
Commodity Computing System
Architectures
(Multi- and Many-core architectures
and accelerators)
Communication and I/O Library
Future Studies
Resource Management and Scheduling Systems for Cloud Computing
(OpenStack Nova, Heat; Slurm)
Virtualization
(Hypervisor and Container)
Locality- and NUMA-aware
Communication
Communication Channels
(SR-IOV, IVShmem, IPC-Shm, CMA)
Fault-Tolerance & Consolidation
(Migration)
QoS-aware

• Redesign MVAPICH2 to make it
virtual machine aware
– SR-IOV shows near to native
performance for inter-node point to
point communication
– IVSHMEM offers shared memory based
data access across co-resident VMs
– Locality Detector: maintains the locality
information of co-resident virtual machines
– Communication Coordinator: selects the
communication channel (SR-IOV, IVSHMEM)
adaptively
Overview of MVAPICH2-Virt with SR-IOV and IVSHMEM
J. Zhang, X. Lu, J. Jose, R. Shi, D. K. Panda. Can Inter-VM
Shmem Benefit MPI Applications on SR-IOV based
Virtualized InfiniBand Clusters? Euro-Par, 2014
J. Zhang, X. Lu, J. Jose, R. Shi, M. Li, D. K. Panda. High
Performance MPI Library over SR-IOV Enabled
InfiniBand Clusters. HiPC, 2014

• OpenStack is one of the most popular
open-source solutions to build clouds and
manage virtual machines
• Deployment with OpenStack
– Supporting SR-IOV configuration
– Supporting IVSHMEM configuration
– Virtual Machine aware design of MVAPICH2
with SR-IOV
• An efficient approach to build HPC Clouds
with MVAPICH2-Virt and OpenStack
MVAPICH2-Virt with SR-IOV and IVSHMEM over OpenStack
J. Zhang, X. Lu, M. Arnold, D. K. Panda. MVAPICH2 over OpenStack with SR-IOV: An Efficient Approach to
Build HPC Clouds. CCGrid, 2015

0
50
100
150
200
250
300
350
400
milc leslie3d pop2 GAPgeofem zeusmp2 lu
ExecutionTime(s)
MV2-SR-IOV-Def
MV2-SR-IOV-Opt
MV2-Native
1%
9.5%
0
1000
2000
3000
4000
5000
6000
22,20 24,10 24,16 24,20 26,10 26,16
ExecutionTime(ms)
Problem Size (Scale, Edgefactor)
MV2-SR-IOV-Def
MV2-SR-IOV-Opt
MV2-Native
2%
• 32 VMs, 6 Core/VM
• Compared to Native, 2-5% overhead for Graph500 with 128 Procs
• Compared to Native, 1-9.5% overhead for SPEC MPI2007 with 128 Procs
Application-Level Performance on Chameleon
SPEC MPI2007Graph500
5%

Execute Live Migration with SR-IOV Device

High Performance SR-IOV enabled VM Migration Support in
MVAPICH2
J. Zhang, X. Lu, D. K. Panda. High-Performance Virtual Machine Migration Framework for MPI Applications on SR-IOV
enabled InfiniBand Clusters. IPDPS, 2017
• Migration with SR-IOV device has to handle the
challenges of detachment/re-attachment of
virtualized IB device and IB connection
• Consist of SR-IOV enabled IB Cluster and External
Migration Controller
• Multiple parallel libraries to notify MPI
applications during migration (detach/reattach
SR-IOV/IVShmem, migrate VMs, migration status)
• Handle the IB connection suspending and
reactivating
• Propose Progress engine (PE) and migration
thread based (MT) design to optimize VM
migration and MPI application performance

• Compared with the TCP, the RDMA scheme reduces the total migration time by 20%
• Total time is dominated by `Migration’ time; Times on other steps are similar across different schemes
• Proposed migration framework could reduce up to 51% migration time
Performance Evaluation of VM Migration Framework
0
0.5
1
1.5
2
2.5
3
TCP IPoIB RDMA
Times(s)
Set POST_MIGRATION Add IVSHMEM Attach VF
Migration Remove IVSHMEM Detach VF
Set PRE_MIGRATION
Breakdown of VM migration
0
5
10
15
20
25
30
35
2 VM 4 VM 8 VM 16 VM
Time(s)
Sequential Migration Framework
Proposed Migration Framework
Multiple VM Migration Time

Bcast (4VMs * 2Procs/VM)
• Migrate a VM from one machine to another while benchmark is running inside
• Proposed MT-based designs perform slightly worse than PE-based designs because of lock/unlock
• No benefit from MT because of NO computation involved
Pt2Pt Latency
0
20
40
60
80
100
120
140
160
180
200
1 4 16 64 256 1K 4K 16K 64K 256K 1M
Latency(us)
Message Size ( bytes)
PE-IPoIB
PE-RDMA
MT-IPoIB
MT-RDMA
0
100
200
300
400
500
600
700
1 4 16 64 256 1K 4K 16K 64K 256K 1M
2Latency(us)
Message Size ( bytes)
PE-IPoIB
PE-RDMA
MT-IPoIB
MT-RDMA

Graph500
• 8 VMs in total and 1 VM carries out migration during application running
• Compared with NM, MT- worst and PE incur some overhead compared with NM
• MT-typical allows migration to be completely overlapped with computation
NAS
0
20
40
60
80
100
120
LU.C EP.C IS.C MG.C CG.C
ExecutionTime(s)
PE
MT-worst
MT-typical
NM
0.0
0.1
0.1
0.2
0.2
0.3
0.3
0.4
0.4
20,10 20,16 20,20 22,10
ExecutionTime(s)
PE
MT-worst
MT-typical
NM

• Container-based technologies (e.g., Docker) provide lightweight virtualization solutions
• Container-based virtualization – share host kernel by containers
Overview of Containers-based Virtualization
VM1
Container1

Benefits of Containers-based Virtualization for HPC on Cloud
65 65.5
178.7
253.7
0
50
100
150
200
250
300
Native-16P 1Conts*16P 2Conts*8P 4Conts*4P
BFSExecutionTime(ms)
Scale, Edgefactor (20,16)
• Experiment on NFS Chameleon Cloud
• Container has less overhead than VM
• BFS time in Graph 500 significantly increases as the number of container increases on one host. Why?
ib_send_lat Graph500
0
1
2
3
4
5
6
7
8
9
1 2 4 8 16 32 64 128 256 512 1K 2K 4K 8K 16K
Latency(us)
Message Size (bytes)
VM-PT
VM-SR-IOV
Container-PT
Native
J. Zhang, X. Lu, D. K. Panda. Performance Characterization of Hypervisor- and Container-Based Virtualization
for HPC on SR-IOV Enabled InfiniBand Clusters. IPDRM, IPDPS Workshop, 2016

• What are the performance bottlenecks when
running MPI applications on multiple
containers per host in HPC cloud?
• Can we propose a new design to overcome the
bottleneck on such container-based HPC
cloud?
• Can optimized design deliver near-native
performance for different container
deployment scenarios?
• Locality-aware based design to enable CMA
and Shared memory channels for MPI
communication across co-resident containers
Containers-based Design: Issues, Challenges, and Approaches
J. Zhang, X. Lu, D. K. Panda. High Performance MPI Library for Container-based HPC Cloud on InfiniBand Clusters.
ICPP, 2016

0
10
20
30
40
50
60
70
80
90
100
MG.D FT.D EP.D LU.D CG.D
ExecutionTime(s)
Container-Def
Container-Opt
Native
• 64 Containers across 16 nodes, pining 4 Cores per Container
• Compared to Container-Def, up to 11% and 73% of execution time reduction for NAS and Graph 500
• Compared to Native, less than 9 % and 5% overhead for NAS and Graph 500
Application-Level Performance on Docker with MVAPICH2
Graph 500
NAS
11%
0
50
100
150
200
250
300
1Cont*16P 2Conts*8P 4Conts*4P
Scale, Edgefactor (20,16)
Container-Def
Container-Opt
Native
73%

• Less than 18% overhead on latency
• Less than 13% overhead on BW
MVAPICH2 Intra-Node and Inter-Node Point-to-Point
Performance on Singularity
0
2
4
6
8
10
12
14
16
18
1 4 16 64 256 1024 4096 16384 65536
Latency(us)
Message Size (Byte)
Latency
Singularity-Intra-Node
Native-Intra-Node
Singularity-Inter-Node
Native-Inter-Node
18%
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
Bandwidth(MB/s)
Message Size (Byte)
BW
Singularity-Intra-Node
Native-Intra-Node
Singularity-Inter-Node
Native-Inter-Node
13%

• 512 Processes across 32 nodes
• Less than 15% and 14% overhead for Bcast and Allreduce, respectively
MVAPICH2 Collective Performance on Singularity
0
10
20
30
40
50
60
70
80
1 4 16 64 256 1024 4096 16384 65536
Latency(us)
Message Size (Byte)
Bcast
Singularity
Native
15%
0
20
40
60
80
100
120
140
160
180
200
4 16 64 256 1024 4096 16384 65536
Latency(us)
Message Size (Byte)
Allreduce
Singularity
Native
14%

• 512 Processes across 32 nodes
• Less than 16% and 11% overhead for NPB and Graph500, respectively
Application-Level Performance on Singularity with MVAPICH2
0
500
1000
1500
2000
2500
3000
22,16 22,20 24,16 24,20 26,16 26,20
Graph500
Singularity
Native
16%
11%
0
50
100
150
200
250
300
CG EP FT IS LU MG
ExecutionTime(s)
NPB (Class D)
Singularity
Native

Nested Virtualization: Containers over Virtual Machines
• Useful for live migration, sandbox application, legacy system
integration, software deployment, etc.
• Performance issues because of the redundant call stacks (two-layer
virtualization) and isolated physical resources

Multiple Communication Paths in Nested Virtualization
1. Intra-VM Intra-Container (across core 4 and core 5)
2. Intra-VM Inter-Container (across core 13 and core 14)
3. Inter-VM Inter-Container (across core 6 and core 12)
4. Inter-Node Inter-Container (across core 15 and the core on remote node)
• Different VM placements introduce multiple communication paths
on container level

Performance Characteristics on Communication Paths
• Two VMs are deployed on the same socket and different sockets, respectively
• *-Def and Inter-VM Inter-Container-1Layer have similar performance
• Large gap compared to native performance
0
10
20
30
40
1 4 16 64 256 1024 4096 16384 66636
Latency(us)
Message Size (Byte)
Intra-Socket
Intra-VM Inter-Container-Def
Inter-VM Inter-Container-Def
Intra-VM Inter-Container-1Layer*
Inter-VM Inter-Container-1Layer*
Native
0
5
10
15
20
1 4 16 64 256 1024 4096 16384 66636
Latency(us)
Message Size (Byte)
Inter-Socket
Intra-VM Inter-Container-Def
Inter-VM Inter-Container-Def
Intra-VM Inter-Container-1Layer*
Inter-VM Inter-Container-1Layer*
Native
1Layer* - J. Zhang, X. Lu, D. K. Panda. High
Performance MPI Library for Container-based HPC
Cloud on InfiniBand, ICPP, 2016

Challenges of Nested Virtualization
• How to further reduce the performance overhead of running applications on
the nested virtualization environment?
• What are the impacts of the different VM/container placement schemes for
the communication on the container level?
• Can we propose a design which can adapt these different VM/container
placement schemes and deliver near-native performance for nested
virtualization environments?

Overview of Proposed Design in MVAPICH2
Two-Layer Locality Detector: Dynamically
detecting MPI processes in the co-
resident containers inside one VM as well
as the ones in the co-resident VMs
Two-Layer NUMA Aware
Communication Coordinator:
Leverage nested locality info, NUMA
architecture info and message to
select appropriate communication
channel
J. Zhang, X. Lu, D. K. Panda. Designing Locality and NUMA Aware MPI Runtime for Nested Virtualization
based HPC Cloud with SR-IOV Enabled InfiniBand, VEE, 2017

Inter-VM Inter-Container Pt2Pt (Intra-Socket)
• 1Layer has similar performance to the Default
• Compared with 1Layer, 2Layer delivers up to 84% and 184% improvement for
latency and BW
Latency BW

Inter-VM Inter-Container Pt2Pt (Inter-Socket)
Latency BW
• 1-Layer has similar performance to the Default
• 2-Layer has near-native performance for small msg, but clear overhead on large msg
• Compared to 2-Layer, Hybrid design brings up to 42% and 25% improvement for
latency and BW, respectively

Application-level Evaluations
• 256 processes across 64 containers on 16 nodes
• Compared with Default, enhanced-hybrid design reduces up to 16% (28,16) and 10% (LU) of
execution time for Graph 500 and NAS, respectively
• Compared with the 1Layer case, enhanced-hybrid design also brings up to 12% (28,16) and 6%
(LU) performance benefit
Class D NASGraph500

• Requirement of managing and isolating virtualized resources of SR-IOV and IVSHMEM
• Such kind of management and isolation is hard to be achieved by MPI library alone, but
much easier with SLURM
• Efficient running MPI applications on HPC Clouds needs SLURM to support managing
SR-IOV and IVSHMEM
– Can critical HPC resources be efficiently shared among users by extending SLURM with
support for SR-IOV and IVSHMEM based virtualization?
– Can SR-IOV and IVSHMEM enabled SLURM and MPI library provide bare-metal performance
for end applications on HPC Clouds?
Need for Supporting SR-IOV and IVSHMEM in SLURM

load SPANK
reclaim VMs
register job
step reply
register job
step req
Slurmctld Slurmd Slurmd
release hosts
run job step req
run job step reply
mpirun_vm
MPI Job
across VMs
VM Config
Reader
load SPANK
VM Launcher
load SPANK
VM Reclaimer
• VM Configuration Reader –
Register all VM configuration
options, set in the job control
environment so that they are
visible to all allocated nodes.
• VM Launcher – Setup VMs on
each allocated nodes.
- File based lock to detect occupied VF
and exclusively allocate free VF
- Assign a unique ID to each IVSHMEM
and dynamically attach to each VM
• VM Reclaimer – Tear down
VMs and reclaim resources
SLURM SPANK Plugin based Design
MPIMPI
vm hostfile

• VM Configuration Reader – VM
options register
• VM Launcher, VM Reclaimer –
Offload to underlying OpenStack
infrastructure
- PCI Whitelist to passthrough free VF to VM
- Extend Nova to enable IVSHMEM when
launching VM
SLURM SPANK Plugin with OpenStack based Design
J. Zhang, X. Lu, S. Chakraborty, D. K. Panda.
SLURM-V: Extending SLURM for Building Efficient
HPC Cloud with SR-IOV and IVShmem. Euro-Par,
2016
reclaim VMs
register job
step reply
register job
step req
Slurmctld Slurmd
release
hosts
launch VM
mpirun_vm
load SPANK
VM Config
Reader
MPI
VM hostfile
OpenStack
daemon
request launch VM
VM Launcher
return
request reclaim VM
VM Reclaimer
return
......
......
......
......

• 32 VMs across 8 nodes, 6 Core/VM
• EASJ - Compared to Native, less than 4% overhead with 128 Procs
• SACJ, EACJ – Also minor overhead, when running NAS as concurrent job with 64 Procs
Application-Level Performance on Chameleon (Graph500)
Exclusive Allocations
Sequential Jobs
0
500
1000
1500
2000
2500
3000
24,16 24,20 26,10
VM
Native
0
50
100
150
200
250
22,10 22,16 22,20BFSExecutionTime(ms)
VM
Native
0
50
100
150
200
250
22 10 22 16 22 20
VM
Native
Shared-host Allocations
Concurrent Jobs
Exclusive Allocations
Concurrent Jobs
4%

• RDMA for Apache Spark
• RDMA for Apache Hadoop 2.x (RDMA-Hadoop-2.x)
– Plugins for Apache, Hortonworks (HDP) and Cloudera (CDH) Hadoop distributions
• RDMA for Apache HBase
• RDMA for Memcached (RDMA-Memcached)
• RDMA for Apache Hadoop 1.x (RDMA-Hadoop)
• OSU HiBD-Benchmarks (OHB)
– HDFS, Memcached, HBase, and Spark Micro-benchmarks
• http://hibd.cse.ohio-state.edu
• Users Base: 215 organizations from 29 countries
• More than 21,000 downloads from the project site
The High-Performance Big Data (HiBD) Project
Available for InfiniBand and RoCE

High-Performance Apache Hadoop over Clouds: Challenges
• How about performance characteristics of native IB-based designs for Apache
Hadoop over SR-IOV enabled cloud environment?
• To achieve locality-aware communication, how can the cluster topology be
automatically detected in a scalable and efficient manner and be exposed to the
Hadoop framework?
• How can we design virtualization-aware policies in Hadoop for efficiently taking
advantage of the detected topology?
• Can the proposed policies improve the performance and fault tolerance of
Hadoop on virtualized platforms?
“How can we design a high-performance Hadoop library for Cloud-based systems?”

Overview of RDMA-Hadoop-Virt Architecture
• Virtualization-aware modules in all the four
main Hadoop components:
– HDFS: Virtualization-aware Block Management
to improve fault-tolerance
– YARN: Extensions to Container Allocation Policy
to reduce network traffic
– MapReduce: Extensions to Map Task Scheduling
Policy to reduce network traffic
– Hadoop Common: Topology Detection Module
for automatic topology detection
• Communications in HDFS, MapReduce, and RPC
go through RDMA-based designs over SR-IOV
enabled InfiniBand
HDFS
YARN
HadoopCommon
MapReduce
HBase Others
Virtual Machines Bare-Metal nodesContainers
Big Data ApplicationsTopologyDetectionModule
Map Task Scheduling
Policy Extension
Container Allocation
Policy Extension
CloudBurst MR-MS Polygraph Others
Virtualization Aware
Block Management
S. Gugnani, X. Lu, D. K. Panda. Designing Virtualization-aware and Automatic Topology Detection Schemes for Accelerating Hadoop on
SR-IOV-enabled Clouds. CloudCom, 2016.

Evaluation with Applications
– 14% and 24% improvement with Default Mode for CloudBurst and Self-Join
– 30% and 55% improvement with Distributed Mode for CloudBurst and Self-Join
0
20
40
60
80
100
Default Mode Distributed Mode
EXECUTIONTIME CloudBurst
RDMA-Hadoop RDMA-Hadoop-Virt
0
50
100
150
200
250
300
350
400
Default Mode Distributed Mode
EXECUTIONTIME
Self-Join
RDMA-Hadoop RDMA-Hadoop-Virt
30%
reduction
55%
reduction

• Distributed Cloud-based Object Storage Service
• Deployed as part of OpenStack installation
• Can be deployed as standalone storage solution as well
• Worldwide data access via Internet
– HTTP-based
• Architecture
– Multiple Object Servers: To store data
– Few Proxy Servers: Act as a proxy for all requests
– Ring: Handles metadata
• Usage
– Input/output source for Big Data applications (most common use
case)
– Software/Data backup
– Storage of VM/Docker images
• Based on traditional TCP sockets communication
OpenStack Swift Overview
Send PUT or GET request
PUT/GET /v1/<account>/<container>/<object>
Proxy
Server
Object
Server
Object
Server
Object
Server
Ring
Disk 1
Disk 2
Disk 1
Disk 2
Disk 1
Disk 2
Swift Architecture

• Challenges
– Proxy server is a bottleneck for large scale deployments
– Object upload/download operations network intensive
– Can an RDMA-based approach benefit?
• Design
– Re-designed Swift architecture for improved scalability and
performance; Two proposed designs:
• Client-Oblivious Design: No changes required on the client side
• Metadata Server-based Design: Direct communication between
client and object servers; bypass proxy server
– RDMA-based communication framework for accelerating
networking performance
– High-performance I/O framework to provide maximum
overlap between communication and I/O
Swift-X: Accelerating OpenStack Swift with RDMA for Building
Efficient HPC Clouds
S. Gugnani, X. Lu, and D. K. Panda, Swift-X: Accelerating OpenStack Swift with RDMA for Building an Efficient HPC Cloud,
accepted at CCGrid’17, May 2017
Client-Oblivious Design
(D1)
Metadata Server-based
Design (D2)

0
5
10
15
20
25
Swift
PUT
Swift-X
(D1)
PUT
Swift-X
(D2)
PUT
Swift
GET
Swift-X
(D1)
GET
Swift-X
(D2)
GET
LATENCY(S)
TIME BREAKUP OF GET AND PUT
Communication I/O Hashsum Other
Swift-X: Accelerating OpenStack Swift with RDMA for Building
Efficient HPC Clouds
0
5
10
15
20
1MB 4MB 16MB 64MB 256MB 1GB 4GB
LATENCY(s)
OBJECT SIZE
GET LATENCY EVALUATION
Swift Swift-X (D2) Swift-X (D1)
Reduced
by 66%
• Up to 66% reduction in GET latency• Communication time reduced by up to
3.8x for PUT and up to 2.8x for GET

Available Appliances on Chameleon Cloud*
Appliance Description
CentOS 7 KVM SR-
IOV
Chameleon bare-metal image customized with the KVM hypervisor and a
recompiled kernel to enable SR-IOV over InfiniBand.
https://www.chameleoncloud.org/appliances/3/
MPI bare-metal
cluster complex
appliance (Based on
Heat)
This appliance deploys an MPI cluster composed of bare metal instances using the
MVAPICH2 library over InfiniBand.
MPI + SR-IOV KVM
cluster (Based on
Heat)
This appliance deploys an MPI cluster of KVM virtual machines using the
MVAPICH2-Virt implementation and configured with SR-IOV for high-performance
communication over InfiniBand. https://www.chameleoncloud.org/appliances/28/
CentOS 7 SR-IOV
RDMA-Hadoop
The CentOS 7 SR-IOV RDMA-Hadoop appliance is built from the CentOS 7
appliance and additionally contains RDMA-Hadoop library with SR-IOV.
• Through these available appliances, users and researchers can easily deploy HPC clouds to perform experiments and run jobs with
– High-Performance SR-IOV + InfiniBand
– High-Performance MVAPICH2 Library over bare-metal InfiniBand clusters
– High-Performance MVAPICH2 Library with Virtualization Support over SR-IOV enabled KVM clusters
– High-Performance Hadoop with RDMA-based Enhancements Support
[*] Only include appliances contributed by OSU NowLab

• MVAPICH2-Virt over SR-IOV-enabled InfiniBand is an efficient approach to build HPC Clouds
– Standalone, OpenStack, Slurm, and Slurm + OpenStack
– Support Virtual Machine Migration with SR-IOV InfiniBand devices
– Support Virtual Machine, Container (Docker and Singularity), and Nested Virtualization
• Very little overhead with virtualization, near native performance at application level
• Much better performance than Amazon EC2
• MVAPICH2-Virt is available for building HPC Clouds
– SR-IOV, IVSHMEM, Docker support, OpenStack
• Big Data analytics stacks such as RDMA-Hadoop can benefit from cloud-aware designs
• Appliances for MVAPICH2-Virt and RDMA-Hadoop are available for building HPC Clouds
• Future releases for supporting running MPI jobs in VMs/Containers with SLURM, etc.
• SR-IOV/container support and appliances for other MVAPICH2 libraries (MVAPICH2-X,
MVAPICH2-GDR, ...) and RDMA-Spark/Memcached
Conclusions

One More Presentation
• Friday (03/31/17) at 11:00am
NVM-aware RDMA-Based Communication and I/O Schemes for High-Perf Big
Data Analytics

Funding Acknowledgments
Funding Support by
Equipment Support by

Personnel Acknowledgments
Current Students
– A. Awan (Ph.D.)
– R. Biswas (M.S.)
– M. Bayatpour (Ph.D.)
– S. Chakraborthy (Ph.D.)
Past Students
– A. Augustine (M.S.)
– P. Balaji (Ph.D.)
– S. Bhagvat (M.S.)
– A. Bhat (M.S.)
– D. Buntinas (Ph.D.)
– L. Chai (Ph.D.)
– B. Chandrasekharan (M.S.)
– N. Dandapanthula (M.S.)
– V. Dhanraj (M.S.)
– T. Gangadharappa (M.S.)
– K. Gopalakrishnan (M.S.)
– R. Rajachandrasekar (Ph.D.)
– G. Santhanaraman (Ph.D.)
– A. Singh (Ph.D.)
– J. Sridhar (M.S.)
– S. Sur (Ph.D.)
– H. Subramoni (Ph.D.)
– K. Vaidyanathan (Ph.D.)
– A. Vishnu (Ph.D.)
– J. Wu (Ph.D.)
– W. Yu (Ph.D.)
Past Research Scientist
– K. Hamidouche
– S. Sur
Past Post-Docs
– D. Banerjee
– X. Besseron
– H.-W. Jin
– W. Huang (Ph.D.)
– W. Jiang (M.S.)
– J. Jose (Ph.D.)
– S. Kini (M.S.)
– M. Koop (Ph.D.)
– K. Kulkarni (M.S.)
– R. Kumar (M.S.)
– S. Krishnamoorthy (M.S.)
– K. Kandalla (Ph.D.)
– P. Lai (M.S.)
– J. Liu (Ph.D.)
– M. Luo (Ph.D.)
– A. Mamidala (Ph.D.)
– G. Marsh (M.S.)
– V. Meshram (M.S.)
– A. Moody (M.S.)
– S. Naravula (Ph.D.)
– R. Noronha (Ph.D.)
– X. Ouyang (Ph.D.)
– S. Pai (M.S.)
– S. Potluri (Ph.D.)
– C.-H. Chu (Ph.D.)
– S. Guganani (Ph.D.)
– J. Hashmi (Ph.D.)
– H. Javed (Ph.D.)
– J. Lin
– M. Luo
– E. Mancini
Current Research Scientists
– X. Lu
– H. Subramoni
Past Programmers
– D. Bureddy
– M. Arnold
– J. Perkins
Current Research Specialist
– J. Smith
– M. Li (Ph.D.)
– D. Shankar (Ph.D.)
– H. Shi (Ph.D.)
– J. Zhang (Ph.D.)
– S. Marcarelli
– J. Vienne
– H. Wang

{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/

Building Efficient HPC Clouds with MCAPICH2 and RDMA-Hadoop over SR-IOV InfiniBand Clusters

Related slideshows

More Related Content

Building Efficient HPC Clouds with MCAPICH2 and RDMA-Hadoop over SR-IOV InfiniBand Clusters