Towards Asynchronous and MPI-Interoperable Active Messages
description
Transcript of Towards Asynchronous and MPI-Interoperable Active Messages
![Page 1: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/1.jpg)
TOWARDS ASYNCHRONOUSAND MPI-INTEROPERABLEACTIVE MESSAGESXin Zhao, Darius Buntinas, Judicael Zounmevo,James Dinan, David Goodell, Pavan Balaji,Rajeev Thakur, Ahmad Afsahi, William Gropp
University of Illinois at Urbana-Champaign,Argonne National Laboratory,Queen’s University
13th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing
![Page 2: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/2.jpg)
Motivation• Many new important large-scale applications• Bioinformatics, social network analysis
2
![Page 3: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/3.jpg)
Motivation• Many new important large-scale applications• Bioinformatics, social network analysis
• Different from “traditional” applications• Many small messages sent to random
targets• Communication pattern is irregular• Computation is data-driven
3
![Page 4: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/4.jpg)
Motivation• Many new important large-scale applications• Bioinformatics, social network analysis
• Different from “traditional” applications• Many small messages sent to random
nodes• Communication pattern is irregular• Computation is data-driven
• Approaches for “traditional” applications are not well-suited
4
![Page 5: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/5.jpg)
while any rank’s queue is not empty: for i in ranks: out_queue[i] empty⟵ for vertex v in my queue: if color(v) is white: color(v) black⟵ for vertex w in neighbors(v): append w to out_queue[owner(w)]
Breath-First Search
for i in ranks: start receiving in_queue[i] from rank ifor j in ranks: start sending out_queue[j] to rank jsynchronize and finish communications
rank 0 rank 1 rank 2
out
in
5
![Page 6: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/6.jpg)
Active Messages (AM)
6
rank 0 rank 1Messages
Messages handler
Reply
Reply handler
• Proposed by von Eicken et al for Split-C in 1992• Sender explicitly sends message• Upon message’s arrival, message handler is triggered
• Receiver is not explicitly involved• A suitable paradigm for data-driven applications• Data is sent immediately• Communication is asynchronous
![Page 7: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/7.jpg)
Breadth-First Search in Active Messageswhile queue is not empty: new_queue empty⟵ begin AM epoch for vertex v in queue: for vertex w in neighbors(v) send AM to owner(w) end AM epoch queue ⟵ new_queue
AM_handler (vertex v) if color(v) is white: color(v) black⟵ insert v to new_queue
rank 0 rank 1 rank 2
7
![Page 8: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/8.jpg)
Motivation• Most applications are MPI-based, but we don’t want to rewrite the entire application• An incremental approach is needed to modify the application
8
![Page 9: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/9.jpg)
Motivation• Most applications are MPI-based, but we don’t want to rewrite the entire application• An incremental approach is needed to modify the application
• This paper tackles the challenge of supporting Active Messages within MPI framework, so that both traditional MPI communication and Active Messages can be simultaneously utilized by applications
9
![Page 10: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/10.jpg)
API DESIGN
![Page 11: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/11.jpg)
Challenges• Active Messages should work correctly with other MPI messages
• Memory consistency semantics• Consistency between two different active messages
11
![Page 12: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/12.jpg)
Challenges• Active Messages should work correctly with other MPI messages
• Memory consistency semantics• Consistency between two different active messages
• Leverage MPI One-sided (RMA) Interface, which already address those challenges
12
![Page 13: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/13.jpg)
MPI RMA Interface
13
• “One-sided” communication• Origin process specifies all communication parameters• Target process exposes memory (“window”) accessed
by other processes• Three basic operations: Put / Get / Accumulate• Accumulate: simple updates on target process
MPI_Put MPI_Get MPI_Accumulate
![Page 14: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/14.jpg)
API Design• Extend MPI_Accumulate to support user function
• User-defined function• MPI_User_function (void *invec, void *inoutvec, int *len,
MPI_Datatype *dtype)
14
![Page 15: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/15.jpg)
API Design• Extend MPI_Accumulate to support user function
• User-defined function• MPI_User_function (void *invec, void *inoutvec, int *len,
MPI_Datatype *dtype)• Operation creation
• MPI_Op_create (MPI_User_function *user_fn, int commute, MPI_Op *user_op)
15
![Page 16: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/16.jpg)
API Design• Extend MPI_Accumulate to support user function
• User-defined function• MPI_User_function (void *invec, void *inoutvec, int *len,
MPI_Datatype *dtype)• Operation creation
• MPI_Op_create (MPI_User_function *user_fn, int commute, MPI_Op *user_op)
• Operation registration• MPIX_Op_register (MPI_Op user_op, int id, MPI_Win win)• Collective call on the window
16
![Page 17: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/17.jpg)
rank 1rank 0
Examplevoid func0 (void *in, void *inout, int *len, MPI_Datatype *dtype);
MPI_Win_create (…, &win);MPI_Op_create (func1, ..., &user_op);MPIX_Op_register (user_op, 0, win);
/* func1 is triggered at rank1 */
MPIX_Op_deregister (user_op, win);MPI_Win_free (&win);
MPI_Win_create (…, &win);MPI_Op_create (func0, ..., &user_op);
MPI_Win_lock (…, 1, …, win);MPI_Accumulate (…, 1, …, user_op, win);MPI_Win_unlock (1, win);
MPIX_Op_deregister (user_op, win);MPI_Win_free (&win);
void func1 (void *in, void *inout, int *len, MPI_Datatype *dtype);
MPIX_Op_register (user_op, 0, win);
func0 and func1 have the same functionality
17
MPIX_Op_register (user_op, 0, win);
![Page 18: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/18.jpg)
ASYNCHRONOUS PROCESSING
![Page 19: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/19.jpg)
Asynchronous execution models• “NON-ASYNC”
• No asynchronous processing, messages are processed by single thread
• Disadvantage: messages cannot be processed upon arrival
19
![Page 20: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/20.jpg)
Asynchronous execution models• “NON-ASYNC”
• No asynchronous processing, messages are processed by single thread
• Disadvantage: messages cannot be processed upon arrival
• “THREAD-ASYNC”• Asynchronous processing is provided by thread on top of
MPI library• Disadvantage: “active polling” for intra-node messages
20
![Page 21: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/21.jpg)
Asynchronous execution models• “NON-ASYNC”
• No asynchronous processing, messages are processed by single thread
• Disadvantage: messages cannot be processed upon arrival
• “THREAD-ASYNC”• Asynchronous processing is provided by thread on top of
MPI library• Disadvantage: “active polling” for intra-node messages
• “INTEGRATED-ASYNC”• Asynchronous processing is supported internally from
MPI library21
![Page 22: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/22.jpg)
Inter-node Communication• Thread-based approach
• Spawn thread in TCP network module
• Block polling for AM• Separate sockets for MPI
messages and AM
spawn AM thread
22
MPIAM
![Page 23: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/23.jpg)
Intra-node Communication• “Origin computation”
• During window creation, processes on the same node allocate a shared-memory region
• Origin process directly fetches data from target process, completes computation locally and writes data back to target process
23
![Page 24: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/24.jpg)
Network
Integrated Idea
rank 0(target)
rank 1(origin)
rank 2(origin)
Shared-memory
NODE 0 NODE 1
24
![Page 25: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/25.jpg)
Network
Integrated IdeaShared-memory
NODE 0 NODE 1
25
rank 0(target)
rank 1(origin)
rank 2(origin)
![Page 26: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/26.jpg)
Network
Integrated IdeaShared-memory
NODE 0 NODE 1
26
rank 0(target)
rank 1(origin)
rank 2(origin)
![Page 27: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/27.jpg)
Network
Integrated IdeaShared-memory
NODE 0 NODE 1
27
rank 0(target)
rank 1(origin)
rank 2(origin)
![Page 28: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/28.jpg)
Network
Integrated IdeaShared-memory
NODE 0 NODE 1
28
rank 0(target)
rank 1(origin)
rank 2(origin)
![Page 29: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/29.jpg)
Network
Integrated IdeaShared-memory
NODE 0 NODE 1
29
rank 0(target)
rank 1(origin)
rank 2(origin)
![Page 30: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/30.jpg)
PERFORMANCE
![Page 31: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/31.jpg)
Stencil Benchmark
31
Number of processes: 2x2 to 20x20, 8 processes per node(“Fusion” Cluster at ANL: 320 nodes, 8 cores per node, QDR Infiniband)
![Page 32: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/32.jpg)
Graph500 Benchmark
32(TEPS: Traversed Edges Per Second, higher is better)
• Breadth-First Search (BFS)• Large number of small messages to random targets• Optimization: origin process accumulates local data and sends only one message to each target process
![Page 33: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/33.jpg)
Conclusion• Why to support Active Messages in MPI• Leverage MPI RMA Interface• API Design• Asynchronous processing
• Intra-node messages• Inter-node messages
• Stencil benchmark and Graph500 benchmark
33
![Page 34: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/34.jpg)
QUESTIONS
![Page 35: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/35.jpg)
BACKUP SLIDES
![Page 36: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/36.jpg)
Message Passing Models• Two-sided communication
• Explicit sends and receives• One-sided communication
• Explicit sends, implicit receives• Simple operations on remote process
• Active Messages• Explicit sends, implicit receives• User-defined operations on remote process
36
![Page 37: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/37.jpg)
Active Message Models• Libraries and runtime systems
• Low level: GASNet, IBM’s LAPI, DCMF and PAMI, AMMPI
• High level: Charm++, X10• Middle: AM++
AM++AMMPI
![Page 38: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/38.jpg)
MPI RMA Interface• Two synchronization modes
Unlock
origin target
ACC(X)epoch
LockPUT(Y)
epoch
Complete
origin target
ACC(X)
StartPUT(Y)
Post
Wait
Passive Target Mode(Lock-Unlock)
Active Target Mode(PSCW, Fence)
![Page 39: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/39.jpg)
Inter-node Communication
39
• Separate sockets for Active Messages
origin_sc
origin_sc rma_sc
rma_scmain thread
AMthread
![Page 40: Towards Asynchronous and MPI-Interoperable Active Messages](https://reader036.fdocuments.net/reader036/viewer/2022062410/568162a7550346895dd32547/html5/thumbnails/40.jpg)
Inter-node Communication• Separate sockets for Active Messages
origin_sc
origin_sc rma_send_sc
rma_send_sc
main thread AM thread
rma_recv_sc
rma_recv_sc
rma_resp_sc
rma_resp_sc