Does your tool support PAPI SDEs yet? 13 th Scalable Tools Workshop - - PowerPoint PPT Presentation

Does your tool support PAPI SDEs yet?

13th Scalable Tools Workshop

Anthony Danalis, Heike Jagode, Jack Dongarra Tahoe City, CA July 28-Aug 1, 2019

Case study: PaRSEC’s task scheduling algorithm

Core 0 Core 0 Core 1 Core 1 Core 2 Core 2 Core N Core N

Core local queues … Shared Global queue (overflow)

Case study: PaRSEC’s task scheduling algorithm

Core 0 Core 0 Core 1 Core 1 Core 2 Core 2 Core N Core N

Core local queues … Shared Global queue (overflow) Thread Local Queues => High Locality Overflow & Work Stealing => Load Balance Shared Global queue (overflow)

Parameter selection

Q1: How long should the local queues be? Q2: Should a thread first steal from a close queue, any queue, or the shared queue?

Parameter selection

Q1: How long should the local queues be? A: 4*Core_Count Q2: Should a thread first steal from a close queue, any queue, or the shared queue? A: Any local queue (closest to farthest), then shared queue.

Testing Benchmark

... ...

...

• 20 Independent Fork-Join chains x 20 (or 25) Tasks per fork.
• Memory bound kernel, with good cache locality.
• 20 Cores on testing node.

...

... ...

...

... ...

...

... ...

...

Execution time vs Local Queue Length

Execution time vs Local Queue Length (zoom)

Execution time vs Local Queue Length (zoom 2)

Execution time vs Local Queue Length (zoom 3)

Execution time vs Local Queue Length (zoom 4)

Execution time vs Local Queue Length (zoom 5)

Execution time vs Local Queue Length (combined)

Failed Stealing Attempts

L2 Cache Misses (L3 show same pattern)

Successful Close Stealing

Successful Close & Far Stealing

Successful Shared Queue Stealing

Successful Local + Shared Queue Stealing

Unanswered questions

Q: So, what causes the bump? Q: How did you measure all these things?

Unanswered questions

Q: So, what causes the bump? A: I don’t know! Q: How did you measure all these things?

Unanswered questions

Q: So, what causes the bump? A: I don’t know! Q: How did you measure all these things? A: I am glad you asked.

What is missing from current infrastructure?

Events that occurred inside the software stack

There is no standardized way for a software layer to export information about its behavior such that other, independently developed, software layers can read it.

HPC Application Math library Task runtime MPI Libibverbs RDMA completion One Sided Communication Data Dependency Distributed Factorization Quantum Chemistry Method

PAPI Software Defined Events

• De facto standard:

SDEs from your library can be read using the standard PAPI_start()/PAPI_stop()/PAPI_read().

• Low overhead:

Performance critical codes can implement SDEs with zero overhead by exporting existing code variables without adding any new instructions in the fast path.

• Rich feature set:

PAPI SDE supports counters, groups, recordings, simple statistics, thread safety, custom callbacks.

The tool infrastructure is already there

The tool infrastructure is already there

Simplest SDE code (library side)

s t a t i c l

• n

g l

• n

g l

• c

a l _ v a r ; v

• i

d s m a l l _ t e s t _ i n i t ( v

• i

d ) { l

• c

a l _ v a r = ; p a p i _ h a n d l e _ t * h a n d l e = p a p i _ s d e _ i n i t ( ” T E S T " ) ; p a p i _ s d e _ r e g i s t e r _ c

• u

n t e r ( h a n d l e , ” E v n t " , P A P I _ S D E _ R O | P A P I _ S D E _ D E L T A , P A P I _ S D E _ l

• n

g _ l

• n

g , & l

• c

a l _ v a r ) ; . . . }

SDE code for registering a callback function

s

• m

e t y p e _ t * d a t a ; v

• i

d s m a l l _ t e s t _ i n i t ( v

• i

d ) { d a t a = . . . p a p i _ h a n d l e _ t * h a n d l e = p a p i _ s d e _ i n i t ( ” T E S T " ) ; p a p i _ s d e _ r e g i s t e r _ f p _ c

• u

n t e r ( h a n d l e , " E v n t " , P A P I _ S D E _ R O | P A P I _ S D E _ D E L T A , P A P I _ S D E _ l

• n

g _ l

• n

g , a c c e s s

• r

, d a t a ) ; . . . }

SDE code for creating a counter (push mode)

v

• i

d * c

• u

n t e r _ h a n d l e ; v

• i

d s m a l l _ t e s t _ i n i t ( v

• i

d ) { p a p i _ h a n d l e _ t * h a n d l e = p a p i _ s d e _ i n i t ( ” T E S T " ) ; p a p i _ s d e _ c r e a t e _ c

• u

n t e r ( h a n d l e , " E v n t " , P A P I _ S D E _ l

• n

g _ l

• n

g , & c

• u

n t e r _ h a n d l e ) ; . . . }

SDE code for creating a recorder (push mode)

v

• i

d * r e c

• r

d e r _ h a n d l e ; v

• i

d s m a l l _ t e s t _ i n i t ( v

• i

d ) { p a p i _ h a n d l e _ t * h a n d l e = p a p i _ s d e _ i n i t ( ” T E S T " ) ; p a p i _ s d e _ c r e a t e _ r e c

• r

d e r ( h a n d l e , " R C R D R " , s i z e

• f

( d

• u

b l e ) , c m p r _ f u n c _ p t r , & r e c

• r

d e r _ h a n d l e ) ; . . . }

SDE code for creating a recorder (push mode)

v

• i

d * r e c

• r

d e r _ h a n d l e ; v

• i

d s m a l l _ t e s t _ i n i t ( v

• i

d ) { p a p i _ h a n d l e _ t * h a n d l e = p a p i _ s d e _ i n i t ( ” T E S T " ) ; p a p i _ s d e _ c r e a t e _ r e c

• r

d e r ( h a n d l e , " R C R D R " , s i z e

• f

( d

• u

b l e ) , c m p r _ f u n c _ p t r , & r e c

• r

d e r _ h a n d l e ) ; . . . }

s d e : : : T E S T : : R C R D R

SDE code for creating a recorder (push mode)

v

• i

d * r e c

• r

d e r _ h a n d l e ; v

• i

d s m a l l _ t e s t _ i n i t ( v

• i

d ) { p a p i _ h a n d l e _ t * h a n d l e = p a p i _ s d e _ i n i t ( ” T E S T " ) ; p a p i _ s d e _ c r e a t e _ r e c

• r

d e r ( h a n d l e , " R C R D R " , s i z e

• f

( d

• u

b l e ) , c m p r _ f u n c _ p t r , & r e c

• r

d e r _ h a n d l e ) ; . . . }

s d e : : : T E S T : : R C R D R s d e : : : T E S T : : R C R D R : C N T

SDE code for creating a recorder (push mode)

v

• i

d * r e c

• r

d e r _ h a n d l e ; v

• i

d s m a l l _ t e s t _ i n i t ( v

• i

d ) { p a p i _ h a n d l e _ t * h a n d l e = p a p i _ s d e _ i n i t ( ” T E S T " ) ; p a p i _ s d e _ c r e a t e _ r e c

• r

d e r ( h a n d l e , " R C R D R " , s i z e

• f

( d

• u

b l e ) , c m p r _ f u n c _ p t r , & r e c

• r

d e r _ h a n d l e ) ; . . . }

s d e : : : T E S T : : R C R D R s d e : : : T E S T : : R C R D R : C N T s d e : : : T E S T : : R C R D R : M I N s d e : : : T E S T : : R C R D R : Q 1 s d e : : : T E S T : : R C R D R : M E D s d e : : : T E S T : : R C R D R : Q 3 s d e : : : T E S T : : R C R D R : M A X

SDE code for updating created counters/recorders

v

• i

d * c

• u

n t e r _ h a n d l e ; v

• i

d * r e c

• r

d e r _ h a n d l e ; v

• i

d p u s h _ t e s t _ d

• w
• r

k ( v

• i

d ) { d

• u

b l e v a l ; l

• n

g l

• n

g i n c r e m e n t = 3 ; v a l = p e r f

• r

m _ u s e f u l _ w

• r

k ( ) ; p a p i _ s d e _ i n c _ c

• u

n t e r ( c

• u

n t e r _ h a n d l e , i n c r e m e n t ) ; p a p i _ s d e _ r e c

• r

d ( r e c

• r

d e r _ h a n d l e , s i z e

• f

( v a l ) , & v a l ) ; }

Performance overheads in simple benchmark

35

Performance overhead in PaRSEC

36

Performance overhead in HPCG

37

Performance overhead in HPCG (zoom)

38

Open Problem for our Community:

What meaningful information to associate with “TASKS_STOLEN”?

– Code location – Hardware events (e.g. cache misses) – Patterns in history (e.g. last task before stealing event) – Patterns in call-path/stack/originating thread

How do we associate useful context information with SDEs?

Conclusions

• Libraries/runtimes generate multiple useful software “events”.
• PAPI SDE allows any software layer to export events.
• SDEs can be read using the standard PAPI functionality.
• SDEs have minimal to zero performance overhead.
• SDEs might require different types of analysis by tools.