P❡r❢♦r♠❛♥❝❡ Pr❡❞✐❝t✐♦♥ ♦❢ ❍P❈ ❆♣♣❧✐❝❛t✐♦♥s❚❤❡ ❙✐♠●r✐❞ Pr♦❥❡❝t

❆✳ ▲❡❣r❛♥❞ ❛♥❞ t❤❡ ❙✐♠●r✐❞ t❡❛♠(M. Quinson, F. Suter, A. Degomme, L. Stanisic, B. Videau,

L. Genovese, S. Thibault, E. Aggulo, . . . )

❈◆❘❙✴■♥r✐❛✴❯♥✐✈❡rs✐t② ♦❢ ●r❡♥♦❜❧❡

▼❛✐s♦♥ ❞❡ ❧❛ s✐♠✉❧❛t✐♦♥▼❛② ✶✵✱ ✷✵✶✻

✶ ✴ ✸✵

❙✉♣❡r❝♦♠♣✉t❡rs

✶✵✵✱✵✵✵ t♦ ✶✱✵✵✵✱✵✵✵ ❝♦r❡s✱ ❛❝❝❡❧❡r❛t♦rs ✭●P❯✱ ❳❡♦♥ P❤✐✮ ❛♥❞ ❛❝♦♠♣❧❡① ❤✐❣❤ s♣❡❡❞ ✐♥t❡r❝♦♥♥❡❝t❆ ✇♦r❧❞✇✐❞❡ ❝♦♠♣❡t✐t✐♦♥ ✭❚♦♣✺✵✵✮

✷ ✴ ✸✵

❈♦♠♣❧❡① ❆♣♣❧✐❝❛t✐♦♥s

▲❛r❣❡r ❛♥❞ ❧❛r❣❡r s❝❛❧❡ ❤②❜r✐❞ ♠❛❝❤✐♥❡s Different programming approaches(e.g., in linear algebra applications)

“Rigid, hand tuned”

SuperLU

Task-based and Dynamic

MUMPS❆♥❛❧②s✐s ❛♥❞ ❈♦♠♣❛r✐s♦♥ ♦❢ ❚✇♦ ❉✐str✐❜✉t❡❞ ▼❡♠♦r② ❙♣❛rs❡ ❙♦❧✈❡rs

❆♠❡st♦②✱ ❉✉✛✱ ▲✬❡①❝❡❧❧❡♥t✱ ▲✐✳ ❆❈▼ ❚r❛♥s✳ ♦♥ ▼❛t❤✳ ❙♦❢t✇❛r❡✱ ❱♦❧✳ ✷✼✱ ◆♦✳ ✹✱ ✷✵✵✶✳

Multitude of technologies

▼P■ ❖♣❡♥▼P ❈✐❧❦ ❈❯❉❆ ❖♣❡♥❈▲ ❈❤❛r♠✰✰

❑❆❆P■ ❙t❛rP❯ ◗❯❆❘❑ P❛r❙❊❈ ❖♠♣❙s ✳✳✳

♥❡❡❞ ❢♦r ♣❡r❢♦r♠❛♥❝❡ ❡✈❛❧✉❛t✐♦♥ ♦♥ ❛ r❡❣✉❧❛r ❜❛s✐s

✸ ✴ ✸✵

❚♦✇❛r❞ ❊①❛s❝❛❧❡ ❄

❆❧r❡❛❞② ✐♥s❛♥❡❧② ❝♦♠♣❧❡① ♣❧❛t❢♦r♠s ❛♥❞ ❛♣♣❧✐❝❛t✐♦♥s ✇✐t❤ P❡t❛✲s❝❛❧❡s②st❡♠s✳ ❉♦ ✇❡ ❤❛✈❡ ❛ ❝❤❛♥❝❡ t♦ ✉♥❞❡rst❛♥❞ ❡①❛s❝❛❧❡ s②st❡♠s ❄

❊✉r♦♣❡❛♥ ❛♣♣r♦❛❝❤ t♦ ❊①❛s❝❛❧❡✿ ▼♦♥t✲❇❧❛♥❝❀ ❧♦✇✲♣♦✇❡r ❝♦♠♠♦❞✐t②❤❛r❞✇❛r❡ s✳❛✳ ❆❘▼✰●P❯s✰❊t❤❡r♥❡t

◆❡❡❞ ❢♦r ❛♣♣❧✐❝❛t✐♦♥ ♣❡r❢♦r♠❛♥❝❡ ♣r❡❞✐❝t✐♦♥ ❛♥❞ ❝❛♣❛❝✐t② ♣❧❛♥♥✐♥❣

▼P■ s✐♠✉❧❛t✐♦♥✿ ✇❤❛t ❢♦r ❄✶ ❍❡❧♣✐♥❣ ❛♣♣❧✐❝❛t✐♦♥ ❞❡✈❡❧♦♣❡rs

Non-intrusive tracing and repeatable executionClassical debugging tools (gdb, valgrind) can be usedSave computing resources (runs on your laptop if possible)

✷ ❍❡❧♣✐♥❣ ❛♣♣❧✐❝❛t✐♦♥ ✉s❡rs

How much resources should I ask for? (scaling)Configure MPI collective operationsProvide baseline

✸ ❈❛♣❛❝✐t② ♣❧❛♥♥✐♥❣ ✭❝❛♥ ✇❡ s❛✈❡ ♦♥ ❝♦♠♣♦♥❡♥ts❄ ✇❤❛t✲✐❢ ❛♥❛❧②s✐s✮

✹ ✴ ✸✵

❋❧♦✉r✐s❤✐♥❣ st❛t❡ ♦❢ t❤❡ ❆rt

❚❤❡r❡ ❛r❡ ♠❛♥② ❞✐✛❡r❡♥t ♣r♦❥❡❝ts✿

❉✐♠❡♠❛s ✭❇❙❈✱ ♣r♦❜❛❜❧② ♦♥❡ ♦❢ t❤❡ ❡❛r❧✐❡st✮P❙■◆❙ ✭❙❉❙❈✱ ✉s❡❞ t♦ r❡❧② ♦♥ ❉✐♠❡♠❛s✮❇✐❣❙✐♠ ✭❯■❯❈✮✿ ❇✐❣◆❡t❙✐♠ ♦r ❇✐❣❋❛st❙✐♠▲♦❣●♦♣❙✐♠ ✭❯■❯❈✴❊❚❍❩✮❙❙❚ ✭❙❛♥❞✐❛ ◆❛t✳ ▲❛❜✳✮✿ ▼✐❝r♦ ♦r ▼❛❝r♦✳ ✳ ✳

❙✐♠●r✐❞✿ ❆ ✶✻ ②❡❛rs ♦❧❞ ♦♣❡♥✲s♦✉r❝❡ ♣r♦❥❡❝t✳ ❈♦❧❧❛❜♦r❛t✐♦♥ ❜❡t✇❡❡♥ ❋r❛♥❝❡✭■◆❘■❆✱ ❈◆❘❙✱ ❯♥✐✈✳ ▲②♦♥✱ ◆❛♥❝②✱ ●r❡♥♦❜❧❡✱ ✳ ✳ ✳ ✮✱ ❯❙❆ ✭❯❈❙❉✱❯✳ ❍❛✇❛✐✐✮✱ ❯❑✱ ❆✉str✐❛ ✭❱✐❡♥♥❛✮✳ ✳ ✳

❤tt♣✿✴✴s✐♠❣r✐❞✳❣❢♦r❣❡✳✐♥r✐❛✳❢r

■♥✐t✐❛❧❧② ❢♦❝✉s❡❞ ♦♥ ●r✐❞ s❡tt✐♥❣s✱ ✇❡ ❛r❣✉❡ t❤❛t t❤❡ s❛♠❡t♦♦❧✴t❡❝❤♥✐q✉❡s ❝❛♥ ❜❡ ✉s❡❞ ❢♦r P✷P✱ ❍P❈ ❛♥❞ ❝❧♦✉❞

✺ ✴ ✸✵

❖✉t❧✐♥❡

✶ ❈♦♥t❡①t

✷ ❙✐♠●r✐❞ ❛♥❞ ▼P■✿ ❙▼P■❙▼P■❙✐♠●r✐❞ ❛♥❞ ❆▼P■

✸ ❙✐♠●r✐❞ ❛♥❞ ❙t❛rP❯Pr✐♥❝✐♣❧❡❉❡♥s❡ ▲✐♥❡❛r ❆❧❣❡❜r❛ ❆♣♣❧✐❝❛t✐♦♥s❙♣❛rs❡ ▲✐♥❡❛r ❆❧❣❡❜r❛ ❆♣♣❧✐❝❛t✐♦♥s

✹ ❈♦♥❝❧✉s✐♦♥

❙▼P■ ✕ ❖✤✐♥❡ ✈s✳ ❖♥❧✐♥❡ ❙✐♠✉❧❛t✐♦♥

Timed Trace[0.001000] 0 compute 1e6 0.01000

[0.010028] 0 send 1 1e6 0.009028

[0.040113] 0 recv 3 1e6 0.030085

[0.010028] 1 recv 0 1e6 0.010028

...

time slice

Visualization

Paje

TRIVA

<?xml version=1.0?><!DOCTYPE platform SYSTEM "simgrid.dtd"><platform version="3"><cluster id="griffon" prefix="griffon-" suffix=".grid5000.fr" radical="1-144" power="286.087kf" bw="125MBps" lat="24us" bb_bw="1.25GBps" bb_lat="0" sharing_policy="FULLDUPLEX" />

Platform DescriptionDownUp DownUp DownUp DownUp

10G1G

1−39 40−74 105−14475−104

13G

10G

Limiter

... ...... ...1.5G

1G

Limiter

DownUp

Simulated Execution Time43.232 seconds

Model the machine of your dreams

mpirun tau, PAPI

Trace once on a

simple cluster

SMPISimulated or Emulated

Computations

Simulated Communications

Time IndependentTrace

0 compute 1e6

0 send 1 1e6

0 recv 3 1e6

1 recv 0 1e6

1 compute 1e6

1 send 2 1e6

2 recv 1 1e6

2 compute 1e6

2 send 3 1e6

3 recv 2 1e6

3 compute 1e6

3 send 0 1e6

Replay the traceas many times as

you want

MPI Application

On-line: simulate/emulate unmodified complex applications

- Possible memory folding and shadow execution- Handles non-deterministic applications

Off-line: trace replay

❖✤✐♥❡ s✐♠✉❧❛t✐♦♥✶ ❖❜t❛✐♥ ❛ t✐♠❡ ✐♥❞❡♣❡♥❞❡♥t tr❛❝❡✷ ❘❡♣❧❛② ✐t ♦♥ t♦♣ ♦❢ ❙✐♠●r✐❞ ❛s ♦❢t❡♥ ❛s ❞❡s✐r❡❞✸ ❆♥❛❧②③❡ ✇✐t❤ t❤❡ ❝♦♠❢♦rt ♦❢ ❛ s✐♠✉❧❛t♦r

❋❛st✱ ❜✉t r❡q✉✐r❡s ❡①tr❛♣♦❧❛t✐♦♥ ❛♥❞ ❧✐♠✐t❡❞ t♦ ♥♦♥✲❛❞❛♣t✐✈❡ ❝♦❞❡s

✻ ✴ ✸✵

❙▼P■ ✕ ❖✤✐♥❡ ✈s✳ ❖♥❧✐♥❡ ❙✐♠✉❧❛t✐♦♥

Timed Trace[0.001000] 0 compute 1e6 0.01000

[0.010028] 0 send 1 1e6 0.009028

[0.040113] 0 recv 3 1e6 0.030085

[0.010028] 1 recv 0 1e6 0.010028

...

time slice

Visualization

Paje

TRIVA

<?xml version=1.0?><!DOCTYPE platform SYSTEM "simgrid.dtd"><platform version="3"><cluster id="griffon" prefix="griffon-" suffix=".grid5000.fr" radical="1-144" power="286.087kf" bw="125MBps" lat="24us" bb_bw="1.25GBps" bb_lat="0" sharing_policy="FULLDUPLEX" />

Platform DescriptionDownUp DownUp DownUp DownUp

10G1G

1−39 40−74 105−14475−104

13G

10G

Limiter

... ...... ...1.5G

1G

Limiter

DownUp

Simulated Execution Time43.232 seconds

Model the machine of your dreams

mpirun tau, PAPI

Trace once on a

simple cluster

SMPISimulated or Emulated

Computations

Simulated Communications

Time IndependentTrace

0 compute 1e6

0 send 1 1e6

0 recv 3 1e6

1 recv 0 1e6

1 compute 1e6

1 send 2 1e6

2 recv 1 1e6

2 compute 1e6

2 send 3 1e6

3 recv 2 1e6

3 compute 1e6

3 send 0 1e6

Replay the traceas many times as

you want

MPI Application

On-line: simulate/emulate unmodified complex applications

- Possible memory folding and shadow execution- Handles non-deterministic applications

Off-line: trace replay

❖♥❧✐♥❡ s✐♠✉❧❛t✐♦♥

❉✐r❡❝t❧② r✉♥ t❤❡ ❝♦❞❡ ♦♥ t♦♣ ♦❢ ❙✐♠●r✐❞✿ s♦✉r❝❡✲t♦✲s♦✉r❝❡ ✭P❤♦tr❛♥✱

❝♦❝❝✐♥❡❧❧❡✱ ❢✷❝✮✱ ●❖❚ ✐♥❥❡❝t✐♦♥✱ ❝♦♠♣✐❧❡r ♣❛ss ❢♦r ❚▲❙✱ ♠♠❛♣ t❤❡ ❞❛t❛ s❡❣♠❡♥t

P♦ss✐❜❧❡ ♠❡♠♦r② s❤❛r✐♥❣ ❜❡t✇❡❡♥ s✐♠✉❧❛t❡❞ ♣r♦❝❡ss❡s ✭r❡❞✉❝❡s♠❡♠♦r② ❢♦♦t♣r✐♥t✮ ❛♥❞ ❦❡r♥❡❧ s❛♠♣❧✐♥❣ ✭r❡❞✉❝❡s s✐♠✉❧❛t✐♦♥ t✐♠❡✮

❈♦♠♣❧✐❡s ✇✐t❤ ♠♦st ♦❢ t❤❡ ▼P■❈❍✸ t❡sts✉✐t❡✱ ❝♦♠♣❛t✐❜❧❡ ✇✐t❤ ♠❛♥② ❈❋✼✼ ❛♥❞ ❋✾✵ ❝♦❞❡s ✭◆❆❙✱ ▲✐♥P❆❈❑✱ ❙✇❡❡♣✸❉✱ ❇✐❣❉❋❚✱ ❙♣❡❝❋❊▼✸❉✮✻ ✴ ✸✵

P❡r❢♦r♠❛♥❝❡ ❊✈❛❧✉❛t✐♦♥ ❚❤r♦✉❣❤ ❋✐♥❡ ●r❛✐♥ ❙✐♠✉❧❛t✐♦♥

P❛❝❦❡t✲❧❡✈❡❧ ♠♦❞❡❧s ❢✉❧❧ s✐♠✉❧❛t✐♦♥ ♦❢ t❤❡ ✇❤♦❧❡ ♣r♦t♦❝♦❧ st❛❝❦ s♦ ❤♦♣❡❢✉❧❧②♣❡r❢❡❝t✱ ❜✉t

❝♦♠♣❧❡① ♠♦❞❡❧s ❤❛r❞ t♦ ✐♥st❛♥t✐❛t❡ ❛♥❞ ✉♥st❛❜❧❡❋❧♦r❡s ▲✉❝✐♦✱ P❛r❡❞❡s✲❋❛rr❡r❛✱ ❏❛♠♠❡❤✱ ❋❧❡✉r②✱ ❘❡❡❞✳ ❖♣♥❡t ♠♦❞❡❧❡r ❛♥❞ ♥s✲✷✿

❈♦♠♣❛r✐♥❣ t❤❡ ❛❝❝✉r❛❝② ♦❢ ♥❡t✇♦r❦ s✐♠✉❧❛t♦rs ❢♦r ♣❛❝❦❡t✲❧❡✈❡❧ ❛♥❛❧②s✐s ✉s✐♥❣ ❛ ♥❡t✇♦r❦

t❡st❜❡❞✳ ❲❙❊❆❙ ❚r❛♥s❛❝t✐♦♥s ♦♥ ❈♦♠♣✉t❡rs ✷✱ ♥♦✳ ✸ ✭✷✵✵✸✮

✐♥❤❡r❡♥t❧② s❧♦✇ ✭♣❛r❛❧❧❡❧✐s♠ ✇♦♥✬t s❛✈❡ ②♦✉ ❤❡r❡✦✮

s♦♠❡t✐♠❡s ✇r♦♥❣❧② ✐♠♣❧❡♠❡♥t❡❞

✇❤♦ ❝❛♥ ✉♥❞❡rst❛♥❞ t❤❡ ♠❛❝r♦s❝♦♣✐❝ ❜❡❤❛✈✐♦r ♦❢ t❤❡ ❛♣♣❧✐❝❛t✐♦♥ ❄

❲❤❡♥ ✇♦r❦✐♥❣ ❛t t❤❡ ❛♣♣❧✐❝❛t✐♦♥ ❧❡✈❡❧✱ t❤❡r❡ ✐s ❛ ♥❡❡❞ ❢♦r s♦♠❡t❤✐♥❣ ♠♦r❡❤✐❣❤ ❧❡✈❡❧ t❤❛t r❡✢❡❝ts t❤❡ ♠❛❝r♦s❝♦♣✐❝ ❝❤❛r❛❝t❡r✐st✐❝s ♦❢ t❤❡ ♠❛❝❤✐♥❡s

✼ ✴ ✸✵

P❡r❢♦r♠❛♥❝❡ ❊✈❛❧✉❛t✐♦♥ ❚❤r♦✉❣❤ ❋✐♥❡ ●r❛✐♥ ❙✐♠✉❧❛t✐♦♥

P❛❝❦❡t✲❧❡✈❡❧ ♠♦❞❡❧s ❢✉❧❧ s✐♠✉❧❛t✐♦♥ ♦❢ t❤❡ ✇❤♦❧❡ ♣r♦t♦❝♦❧ st❛❝❦ s♦ ❤♦♣❡❢✉❧❧②♣❡r❢❡❝t✱ ❜✉t

❝♦♠♣❧❡① ♠♦❞❡❧s ❤❛r❞ t♦ ✐♥st❛♥t✐❛t❡ ❛♥❞ ✉♥st❛❜❧❡❋❧♦r❡s ▲✉❝✐♦✱ P❛r❡❞❡s✲❋❛rr❡r❛✱ ❏❛♠♠❡❤✱ ❋❧❡✉r②✱ ❘❡❡❞✳ ❖♣♥❡t ♠♦❞❡❧❡r ❛♥❞ ♥s✲✷✿

❈♦♠♣❛r✐♥❣ t❤❡ ❛❝❝✉r❛❝② ♦❢ ♥❡t✇♦r❦ s✐♠✉❧❛t♦rs ❢♦r ♣❛❝❦❡t✲❧❡✈❡❧ ❛♥❛❧②s✐s ✉s✐♥❣ ❛ ♥❡t✇♦r❦

t❡st❜❡❞✳ ❲❙❊❆❙ ❚r❛♥s❛❝t✐♦♥s ♦♥ ❈♦♠♣✉t❡rs ✷✱ ♥♦✳ ✸ ✭✷✵✵✸✮

✐♥❤❡r❡♥t❧② s❧♦✇ ✭♣❛r❛❧❧❡❧✐s♠ ✇♦♥✬t s❛✈❡ ②♦✉ ❤❡r❡✦✮

s♦♠❡t✐♠❡s ✇r♦♥❣❧② ✐♠♣❧❡♠❡♥t❡❞

✇❤♦ ❝❛♥ ✉♥❞❡rst❛♥❞ t❤❡ ♠❛❝r♦s❝♦♣✐❝ ❜❡❤❛✈✐♦r ♦❢ t❤❡ ❛♣♣❧✐❝❛t✐♦♥ ❄

❲❤❡♥ ✇♦r❦✐♥❣ ❛t t❤❡ ❛♣♣❧✐❝❛t✐♦♥ ❧❡✈❡❧✱ t❤❡r❡ ✐s ❛ ♥❡❡❞ ❢♦r s♦♠❡t❤✐♥❣ ♠♦r❡❤✐❣❤ ❧❡✈❡❧ t❤❛t r❡✢❡❝ts t❤❡ ♠❛❝r♦s❝♦♣✐❝ ❝❤❛r❛❝t❡r✐st✐❝s ♦❢ t❤❡ ♠❛❝❤✐♥❡s

✼ ✴ ✸✵

▲♦❣●P❙ ✐♥ ❛ ◆✉ts❤❡❧❧

❚❤❡ ▲♦❣P ♠♦❞❡❧ ✇❛s ✐♥✐t✐❛❧❧② ❞❡s✐❣♥❡❞ ❢♦r ❝♦♠♣❧❡①✐t② ❛♥❛❧②s✐s ❛♥❞❛❧❣♦r✐t❤♠ ❞❡s✐❣♥✳ ▼❛♥② ✈❛r✐❛t✐♦♥s ❛✈❛✐❧❛❜❧❡ t♦ ❛❝❝♦✉♥t ❢♦r ♣r♦t♦❝♦❧ s✇✐t❝❤

Pr

Ps

T✶ T✷ T✸ T✷ T✸T✺ T✶T✹

Ps

Pr

ts

tr

❆s②♥❝❤r♦♥♦✉s ♠♦❞❡ ✭k 6 S ✮ ❘❡♥❞❡③✲✈♦✉s ♠♦❞❡ ✭k > S✮

❚❤❡ Ti ✬s ❛r❡ ❜❛s✐❝❛❧❧② ❝♦♥t✐♥✉♦✉s ❧✐♥❡❛r ❢✉♥❝t✐♦♥s✳

T✶ = o+kOs T✷ =

{

L+ kg ✐❢ k < s

L+ sg + (k − s)G ♦t❤❡r✇✐s❡

T✸ = o+kOr T✹ = ♠❛①(L+o, tr−ts)+o T✺ = ✷o + L

✽ ✴ ✸✵

▲♦❣●P❙ ✐♥ ❛ ◆✉ts❤❡❧❧

❚❤❡ ▲♦❣P ♠♦❞❡❧ ✇❛s ✐♥✐t✐❛❧❧② ❞❡s✐❣♥❡❞ ❢♦r ❝♦♠♣❧❡①✐t② ❛♥❛❧②s✐s ❛♥❞❛❧❣♦r✐t❤♠ ❞❡s✐❣♥✳ ▼❛♥② ✈❛r✐❛t✐♦♥s ❛✈❛✐❧❛❜❧❡ t♦ ❛❝❝♦✉♥t ❢♦r ♣r♦t♦❝♦❧ s✇✐t❝❤

Pr

Ps

T✶ T✷ T✸ T✷ T✸T✺ T✶T✹

Ps

Pr

ts

tr

❆s②♥❝❤r♦♥♦✉s ♠♦❞❡ ✭k 6 S ✮ ❘❡♥❞❡③✲✈♦✉s ♠♦❞❡ ✭k > S✮

❚❤❡ Ti ✬s ❛r❡ ❜❛s✐❝❛❧❧② ❝♦♥t✐♥✉♦✉s ❧✐♥❡❛r ❢✉♥❝t✐♦♥s✳❘♦✉t✐♥❡ ❈♦♥❞✐t✐♦♥ ❈♦st

▼P■❴❙❡♥❞ k 6 S T✶

k > S T✹ + T✺ + T✶

▼P■❴❘❡❝✈ k 6 S ♠❛①(T✶ + T✷ − (tr − ts), ✵) + T✸

k > S ♠❛①(o + L− (tr − ts), ✵) + o+T✺ + T✶ + T✷ + T✸

▼P■❴■s❡♥❞ o

▼P■❴■r❡❝✈ o

✽ ✴ ✸✵

▲♦❣●P❙ ✐♥ ❛ ◆✉ts❤❡❧❧

❚❤❡ ▲♦❣P ♠♦❞❡❧ ✇❛s ✐♥✐t✐❛❧❧② ❞❡s✐❣♥❡❞ ❢♦r ❝♦♠♣❧❡①✐t② ❛♥❛❧②s✐s ❛♥❞❛❧❣♦r✐t❤♠ ❞❡s✐❣♥✳ ▼❛♥② ✈❛r✐❛t✐♦♥s ❛✈❛✐❧❛❜❧❡ t♦ ❛❝❝♦✉♥t ❢♦r ♣r♦t♦❝♦❧ s✇✐t❝❤

Pr

Ps

T✶ T✷ T✸ T✷ T✸T✺ T✶T✹

Ps

Pr

ts

tr

❆s②♥❝❤r♦♥♦✉s ♠♦❞❡ ✭k 6 S ✮ ❘❡♥❞❡③✲✈♦✉s ♠♦❞❡ ✭k > S✮

❚❤❡ Ti ✬s ❛r❡ ❜❛s✐❝❛❧❧② ❝♦♥t✐♥✉♦✉s ❧✐♥❡❛r ❢✉♥❝t✐♦♥s✳

▼❛② r❡✢❡❝t t❤❡ ♦♣❡r❛t✐♦♥ ♦❢ s♣❡❝✐❛❧✐③❡❞ ❍P❈ ♥❡t✇♦r❦s ❢r♦♠ t❤❡ ❡❛r❧②✶✾✾✵s✳ ✳ ✳

■❣♥♦r❡s ♣♦t❡♥t✐❛❧❧② ❝♦♥❢♦✉♥❞✐♥❣ ❢❛❝t♦rs ♣r❡s❡♥t ✐♥ ♠♦❞❡r♥✲❞❛② s②st❡♠s✭❡✳❣✳✱ ❝♦♥t❡♥t✐♦♥✱ t♦♣♦❧♦❣②✱ ❝♦♠♣❧❡① ♣r♦t♦❝♦❧ st❛❝❦✱ ✳ ✳ ✳ ✮

❯♥❧❡ss ②♦✉ ❤❛✈❡ ❛ ✇❡❧❧✲t✉♥❡❞ ❤✐❣❤✲❡♥❞ ♠❛❝❤✐♥❡✱ s✉❝❤ ♠♦❞❡❧ ✐s ✉♥❧✐❦❡❧②t♦ ♣r♦✈✐❞❡ ❛❝❝✉r❛t❡ ❡st✐♠❛t✐♦♥s ♦r ✉s❡❢✉❧ ❜❛s❡❧✐♥❡ ❝♦♠♣❛r✐s♦♥s

✽ ✴ ✸✵

▼P■ P♦✐♥t✲t♦✲P♦✐♥t ❈♦♠♠✉♥✐❝❛t✐♦♥

❘❛♥❞♦♠✐③❡❞ ♠❡❛s✉r❡♠❡♥ts ✭❖♣❡♥▼P■✴❚❈P✴❊t❤✶●❇✮ s✐♥❝❡ ✇❡ ❛r❡ ♥♦t✐♥t❡r❡st❡❞ ✐♥ ♣❡❛❦ ♣❡r❢♦r♠❛♥❝❡s ❜✉t ✐♥ ♣❡r❢♦r♠❛♥❝❡ ❝❤❛r❛❝t❡r✐③❛t✐♦♥

Small

Medium1Medium2

Detached

Small

Medium1Medium2

Detached

MPI_Send MPI_Recv

1e−04

1e−02

1e+01 1e+03 1e+05 1e+01 1e+03 1e+05Message size (bytes)

Dur

atio

n (s

econ

ds) group

Small

Medium1

Medium2

Detached

Large

❚❤❡r❡ ✐s ❛ q✉✐t❡ ✐♠♣♦rt❛♥t ✈❛r✐❛❜✐❧✐t②

❚❤❡r❡ ❛r❡ ❛t ❧❡❛st ✹ ❞✐✛❡r❡♥t ♠♦❞❡s

■t ✐s ♣✐❡❝❡✲✇✐s❡ ❧✐♥❡❛r ❛♥❞ ❞✐s❝♦♥t✐♥✉♦✉s

✾ ✴ ✸✵

❙▼P■ ✕ ❍②❜r✐❞ ▼♦❞❡❧

❙▼P■ ❝♦♠❜✐♥❡s ❛❝❝✉r❛t❡ ❞❡s❝r✐♣t✐♦♥ ♦❢ t❤❡ ♣❧❛t❢♦r♠✱ ✇✐t❤ ❜♦t❤ ✢✉✐❞ ❛♥❞▲♦❣P ❢❛♠✐❧② ♠♦❞❡❧s✿

▲♦❣P✿ ♠❡❛s✉r❡ ♦♥ r❡❛❧ ♥♦❞❡s t♦ ❛❝❝✉r❛t❡❧② ♠♦❞❡❧ ♣t✷♣t ♣❡r❢♦r♠❛♥❝❡✭❞✐s❝♦♥t✐♥✉✐t✐❡s✮ ❛♥❞ ❝♦♠♠✉♥✐❝❛t✐♦♥ ♠♦❞❡s ✭❛s②♥❝❤r♦♥♦✉s✱ ❞❡t❛❝❤❡❞✱s②♥❝❤r♦♥♦✉s✮

Ps

Pr

T2T4

T1

Ps

Pr

T4 T2

Pr

Ps

T1 T2 T3

❋❧✉✐❞ ♠♦❞❡❧ ✭❜❛♥❞✇✐❞t❤ s❤❛r✐♥❣✮✿ ❛❝❝♦✉♥t ❢♦r ❝♦♥t❡♥t✐♦♥ ❛♥❞ ♥❡t✇♦r❦t♦♣♦❧♦❣②

1-39 40-74 105-14475-104

1G

10G

DownUp DownUp DownUp DownUp

10G1G

1−39 40−74 105−14475−104

13G

10G

Limiter

... ...... ...1.5G

1G

Limiter

DownUp

✶✵ ✴ ✸✵

❈♦❧❧❡❝t✐✈❡ ❈♦♠♠✉♥✐❝❛t✐♦♥s

❈❧❛ss✐❝❛❧ ❛♣♣r♦❛❝❤❡s✿

✉s❡ s✐♠♣❧❡ ❛♥❛❧②t✐❝❛❧ ❢♦r♠✉❧❛s

❜❡♥❝❤♠❛r❦ ❡✈❡r②t❤✐♥❣ ❛♥❞ ✐♥❥❡❝t ❝♦rr❡s♣♦♥❞✐♥❣ t✐♠✐♥❣

tr❛❝❡ ❝♦♠♠✉♥✐❝❛t✐♦♥ ♣❛tt❡r♥ ❛♥❞ r❡♣❧❛②

❘❡❛❧ ▼P■ ✐♠♣❧❡♠❡♥t❛t✐♦♥s ❤❛✈❡ s❡✈❡r❛❧ ✐♠♣❧❡♠❡♥t❛t✐♦♥s ❢♦r ❡❛❝❤ ❝♦❧❧❡❝t✐✈❡❛♥❞ s❡❧❡❝t t❤❡ r✐❣❤t ♦♥❡ ❛t r✉♥t✐♠❡

✷✸✵✵ ❧✐♥❡s ♦❢ ❝♦❞❡ ❢♦r t❤❡ ❆❧❧❘❡❞✉❝❡ ✐♥ ❖♣❡♥▼P■✦✦✦

❙▼P■ ♥♦✇ ✉s❡s

♠♦r❡ t❤❛♥ ✶✵✵ ❝♦❧❧❡❝t✐✈❡ ❛❧❣♦r✐t❤♠s ❢r♦♠ t❤r❡❡❡①✐st✐♥❣ ✐♠♣❧❡♠❡♥t❛t✐♦♥s ✭▼P■❈❍✱ ❖♣❡♥▼P■✱❙❚❆❘✲▼P■✮ ❝❛♥ ❜❡ s❡❧❡❝t❡❞

t❤❡ s❛♠❡ s❡❧❡❝t✐♦♥ ❧♦❣✐❝ ❛s ▼P■❈❍ ♦r ❖♣❡♥▼P■t♦ ❛❝❝✉r❛t❡❧② s✐♠✉❧❛t❡ t❤❡✐r ❜❡❤❛✈✐♦r

❙✉❝❤ ❛❝❝✉r❛t❡ ♠♦❞❡❧✐♥❣ ✐s ❛❝t✉❛❧❧② ❝r✐t✐❝❛❧ t♦ ♦❜t❛✐♥ ❞❡❝❡♥t ♣r❡❞✐❝t✐♦♥s

✶✶ ✴ ✸✵

❈♦❧❧❡❝t✐✈❡ ❈♦♠♠✉♥✐❝❛t✐♦♥s

❈❧❛ss✐❝❛❧ ❛♣♣r♦❛❝❤❡s✿

✉s❡ s✐♠♣❧❡ ❛♥❛❧②t✐❝❛❧ ❢♦r♠✉❧❛s

❜❡♥❝❤♠❛r❦ ❡✈❡r②t❤✐♥❣ ❛♥❞ ✐♥❥❡❝t ❝♦rr❡s♣♦♥❞✐♥❣ t✐♠✐♥❣

tr❛❝❡ ❝♦♠♠✉♥✐❝❛t✐♦♥ ♣❛tt❡r♥ ❛♥❞ r❡♣❧❛②

❘❡❛❧ ▼P■ ✐♠♣❧❡♠❡♥t❛t✐♦♥s ❤❛✈❡ s❡✈❡r❛❧ ✐♠♣❧❡♠❡♥t❛t✐♦♥s ❢♦r ❡❛❝❤ ❝♦❧❧❡❝t✐✈❡❛♥❞ s❡❧❡❝t t❤❡ r✐❣❤t ♦♥❡ ❛t r✉♥t✐♠❡

✷✸✵✵ ❧✐♥❡s ♦❢ ❝♦❞❡ ❢♦r t❤❡ ❆❧❧❘❡❞✉❝❡ ✐♥ ❖♣❡♥▼P■✦✦✦

❙▼P■ ♥♦✇ ✉s❡s

♠♦r❡ t❤❛♥ ✶✵✵ ❝♦❧❧❡❝t✐✈❡ ❛❧❣♦r✐t❤♠s ❢r♦♠ t❤r❡❡❡①✐st✐♥❣ ✐♠♣❧❡♠❡♥t❛t✐♦♥s ✭▼P■❈❍✱ ❖♣❡♥▼P■✱❙❚❆❘✲▼P■✮ ❝❛♥ ❜❡ s❡❧❡❝t❡❞

t❤❡ s❛♠❡ s❡❧❡❝t✐♦♥ ❧♦❣✐❝ ❛s ▼P■❈❍ ♦r ❖♣❡♥▼P■t♦ ❛❝❝✉r❛t❡❧② s✐♠✉❧❛t❡ t❤❡✐r ❜❡❤❛✈✐♦r

❙✉❝❤ ❛❝❝✉r❛t❡ ♠♦❞❡❧✐♥❣ ✐s ❛❝t✉❛❧❧② ❝r✐t✐❝❛❧ t♦ ♦❜t❛✐♥ ❞❡❝❡♥t ♣r❡❞✐❝t✐♦♥s✶✶ ✴ ✸✵

❱❛❧✐❞❛t✐♦♥✿ ◆♦♥✲tr✐✈✐❛❧ ❆♣♣❧✐❝❛t✐♦♥ ❙❝❛❧✐♥❣ ✭✶✮

❊①♣❡r✐♠❡♥ts r✉♥ ✇✐t❤ s❡✈❡r❛❧ ◆❆❙ ♣❛r❛❧❧❡❧ ❜❡♥❝❤♠❛r❦s t♦ ✭✐♥✮✈❛❧✐❞❛t❡ t❤❡♠♦❞❡❧ ❢♦r ❚❈P ♣❧❛t❢♦r♠

◆♦♥ tr✐✈✐❛❧ s❝❛❧✐♥❣

❱❡r② ❣♦♦❞ ❛❝❝✉r❛❝② ✭❡s♣❡❝✐❛❧❧② ❝♦♠♣❛r❡❞ t♦ ▲♦❣P✮

✉♥❧❡ss ❝♦♥t❡♥t✐♦♥ ❞r✐✈❡s ❚❈P ✐♥ ❛ ❝r❛③② st❛t❡✳ ✳ ✳

B C

qq

qqq

q

qq

q

q

10

15

20

25

30

CG

6432 128168 6432 128168

Number of nodes

Sp

ee

du

p

Model

q Real

LogGPS

Network collapse

SMPI

▼❛ss✐✈❡ s✇✐t❝❤ ♣❛❝❦❡t ❞r♦♣s ❧❡❛❞ t♦ ✷✵✵♠s t✐♠❡♦✉ts ✐♥ ❚❈P✦❚❤✐s ✐s ❛ s♦❢t✇❛r❡ ✐ss✉❡ t❤❛t ♥❡❡❞s t♦ ❜❡ ✜①❡❞ ✭♥♦t ♠♦❞❡❧❡❞✮ ✐♥ r❡❛❧✐t②

✶✷ ✴ ✸✵

❱❛❧✐❞❛t✐♦♥✿ ◆♦♥✲tr✐✈✐❛❧ ❆♣♣❧✐❝❛t✐♦♥ ❙❝❛❧✐♥❣ ✭✶✮

❊①♣❡r✐♠❡♥ts r✉♥ ✇✐t❤ s❡✈❡r❛❧ ◆❆❙ ♣❛r❛❧❧❡❧ ❜❡♥❝❤♠❛r❦s t♦ ✭✐♥✮✈❛❧✐❞❛t❡ t❤❡♠♦❞❡❧ ❢♦r ❚❈P ♣❧❛t❢♦r♠

◆♦♥ tr✐✈✐❛❧ s❝❛❧✐♥❣

❱❡r② ❣♦♦❞ ❛❝❝✉r❛❝② ✭❡s♣❡❝✐❛❧❧② ❝♦♠♣❛r❡❞ t♦ ▲♦❣P✮

✉♥❧❡ss ❝♦♥t❡♥t✐♦♥ ❞r✐✈❡s ❚❈P ✐♥ ❛ ❝r❛③② st❛t❡✳ ✳ ✳

▼❛ss✐✈❡ s✇✐t❝❤ ♣❛❝❦❡t ❞r♦♣s ❧❡❛❞ t♦ ✷✵✵♠s t✐♠❡♦✉ts ✐♥ ❚❈P✦

❚❤✐s ✐s ❛ s♦❢t✇❛r❡ ✐ss✉❡ t❤❛t ♥❡❡❞s t♦ ❜❡ ✜①❡❞ ✭♥♦t ♠♦❞❡❧❡❞✮ ✐♥ r❡❛❧✐t②

✶✷ ✴ ✸✵

❱❛❧✐❞❛t✐♦♥✿ ◆♦♥✲tr✐✈✐❛❧ ❆♣♣❧✐❝❛t✐♦♥ ❙❝❛❧✐♥❣ ✭✶✮

❊①♣❡r✐♠❡♥ts r✉♥ ✇✐t❤ s❡✈❡r❛❧ ◆❆❙ ♣❛r❛❧❧❡❧ ❜❡♥❝❤♠❛r❦s t♦ ✭✐♥✮✈❛❧✐❞❛t❡ t❤❡♠♦❞❡❧ ❢♦r ❚❈P ♣❧❛t❢♦r♠

◆♦♥ tr✐✈✐❛❧ s❝❛❧✐♥❣

❱❡r② ❣♦♦❞ ❛❝❝✉r❛❝② ✭❡s♣❡❝✐❛❧❧② ❝♦♠♣❛r❡❞ t♦ ▲♦❣P✮

✉♥❧❡ss ❝♦♥t❡♥t✐♦♥ ❞r✐✈❡s ❚❈P ✐♥ ❛ ❝r❛③② st❛t❡✳ ✳ ✳

▼❛ss✐✈❡ s✇✐t❝❤ ♣❛❝❦❡t ❞r♦♣s ❧❡❛❞ t♦ ✷✵✵♠s t✐♠❡♦✉ts ✐♥ ❚❈P✦

❚❤✐s ✐s ❛ s♦❢t✇❛r❡ ✐ss✉❡ t❤❛t ♥❡❡❞s t♦ ❜❡ ✜①❡❞ ✭♥♦t ♠♦❞❡❧❡❞✮ ✐♥ r❡❛❧✐t②

✶✷ ✴ ✸✵

❱❛❧✐❞❛t✐♦♥✿ ◆♦♥✲tr✐✈✐❛❧ ❆♣♣❧✐❝❛t✐♦♥ ❙❝❛❧✐♥❣ ✭✶✮

❊①♣❡r✐♠❡♥ts r✉♥ ✇✐t❤ s❡✈❡r❛❧ ◆❆❙ ♣❛r❛❧❧❡❧ ❜❡♥❝❤♠❛r❦s t♦ ✭✐♥✮✈❛❧✐❞❛t❡ t❤❡♠♦❞❡❧ ❢♦r ❚❈P ♣❧❛t❢♦r♠

◆♦♥ tr✐✈✐❛❧ s❝❛❧✐♥❣

❱❡r② ❣♦♦❞ ❛❝❝✉r❛❝② ✭❡s♣❡❝✐❛❧❧② ❝♦♠♣❛r❡❞ t♦ ▲♦❣P✮

✉♥❧❡ss ❝♦♥t❡♥t✐♦♥ ❞r✐✈❡s ❚❈P ✐♥ ❛ ❝r❛③② st❛t❡✳ ✳ ✳

▼❛ss✐✈❡ s✇✐t❝❤ ♣❛❝❦❡t ❞r♦♣s ❧❡❛❞ t♦ ✷✵✵♠s t✐♠❡♦✉ts ✐♥ ❚❈P✦

❚❤✐s ✐s ❛ s♦❢t✇❛r❡ ✐ss✉❡ t❤❛t ♥❡❡❞s t♦ ❜❡ ✜①❡❞ ✭♥♦t ♠♦❞❡❧❡❞✮ ✐♥ r❡❛❧✐t②

✶✷ ✴ ✸✵

❱❛❧✐❞❛t✐♦♥✿ ◆♦♥✲tr✐✈✐❛❧ ❆♣♣❧✐❝❛t✐♦♥ ❙❝❛❧✐♥❣ ✭✶✮

❊①♣❡r✐♠❡♥ts r✉♥ ✇✐t❤ s❡✈❡r❛❧ ◆❆❙ ♣❛r❛❧❧❡❧ ❜❡♥❝❤♠❛r❦s t♦ ✭✐♥✮✈❛❧✐❞❛t❡ t❤❡♠♦❞❡❧ ❢♦r ❚❈P ♣❧❛t❢♦r♠

◆♦♥ tr✐✈✐❛❧ s❝❛❧✐♥❣

❱❡r② ❣♦♦❞ ❛❝❝✉r❛❝② ✭❡s♣❡❝✐❛❧❧② ❝♦♠♣❛r❡❞ t♦ ▲♦❣P✮

✉♥❧❡ss ❝♦♥t❡♥t✐♦♥ ❞r✐✈❡s ❚❈P ✐♥ ❛ ❝r❛③② st❛t❡✳ ✳ ✳B C

qq

qqq

q

qq

q

q

10

15

20

25

30

CG

6432 128168 6432 128168

Number of nodes

Sp

ee

du

p

Model

q Real

LogGPS

Network collapse

SMPI

▼❛ss✐✈❡ s✇✐t❝❤ ♣❛❝❦❡t ❞r♦♣s ❧❡❛❞ t♦ ✷✵✵♠s t✐♠❡♦✉ts ✐♥ ❚❈P✦

❚❤✐s ✐s ❛ s♦❢t✇❛r❡ ✐ss✉❡ t❤❛t ♥❡❡❞s t♦ ❜❡ ✜①❡❞ ✭♥♦t ♠♦❞❡❧❡❞✮ ✐♥ r❡❛❧✐t②

✶✷ ✴ ✸✵

❱❛❧✐❞❛t✐♦♥✿ ◆♦♥✲tr✐✈✐❛❧ ❆♣♣❧✐❝❛t✐♦♥ ❙❝❛❧✐♥❣ ✭✷✮

❊①♣❡r✐♠❡♥ts ❛❧s♦ r✉♥ ✉s✐♥❣ r❡❛❧ P❤②s✐❝s ❝♦❞❡ ✭❇✐❣❉❋❚✱ ❙P❊❈❋❊▼✸❉✮ ♦♥❚✐❜✐❞❛❜♦ ✭❆❘▼ ❝❧✉st❡r ♣r♦t♦t②♣❡✮

❚❤❡ s❡t ♦❢ ❝♦❧❧❡❝t✐✈❡ ♦♣❡r❛t✐♦♥s ♠❛② ❝♦♠♣❧❡t❡❧② ❝❤❛♥❣❡ ❞❡♣❡♥❞✐♥❣ ♦♥t❤❡ ✐♥st❛♥❝❡✱ ❤❡♥❝❡ t❤❡ ♥❡❡❞ t♦ ✉s❡ ♦♥❧✐♥❡ s✐♠✉❧❛t✐♦♥❱❡r② ❣♦♦❞ ❛❝❝✉r❛❝② ✭❡s♣❡❝✐❛❧❧② ❝♦♠♣❛r❡❞ t♦ ▲♦❣P✮

Tibidabo

q

q

q

q

q

50

100

Sm

all

8 16 32 64 128

Number of nodes

Sp

ee

du

p

Model

q Real

Hybrid

LogGPS

Fluid

✶✸ ✴ ✸✵

❖✉t❧✐♥❡

✶ ❈♦♥t❡①t

✷ ❙✐♠●r✐❞ ❛♥❞ ▼P■✿ ❙▼P■❙▼P■❙✐♠●r✐❞ ❛♥❞ ❆▼P■

✸ ❙✐♠●r✐❞ ❛♥❞ ❙t❛rP❯Pr✐♥❝✐♣❧❡❉❡♥s❡ ▲✐♥❡❛r ❆❧❣❡❜r❛ ❆♣♣❧✐❝❛t✐♦♥s❙♣❛rs❡ ▲✐♥❡❛r ❆❧❣❡❜r❛ ❆♣♣❧✐❝❛t✐♦♥s

✹ ❈♦♥❝❧✉s✐♦♥

❘❡♠❡♠❜❡r t❤❡ ❈♦♠♣❧❡① ❆♣♣❧✐❝❛t✐♦♥s ❄

▲❛r❣❡r ❛♥❞ ❧❛r❣❡r s❝❛❧❡ ❤②❜r✐❞ ♠❛❝❤✐♥❡s Different programming approaches(e.g., in linear algebra applications)

“Rigid, hand tuned”

SuperLU

Task-based and Dynamic

MUMPS❆♥❛❧②s✐s ❛♥❞ ❈♦♠♣❛r✐s♦♥ ♦❢ ❚✇♦ ❉✐str✐❜✉t❡❞ ▼❡♠♦r② ❙♣❛rs❡ ❙♦❧✈❡rs

❆♠❡st♦②✱ ❉✉✛✱ ▲✬❡①❝❡❧❧❡♥t✱ ▲✐✳ ❆❈▼ ❚r❛♥s✳ ♦♥ ▼❛t❤✳ ❙♦❢t✇❛r❡✱ ❱♦❧✳ ✷✼✱ ◆♦✳ ✹✱ ✷✵✵✶✳

Multitude of technologies

▼P■ ❖♣❡♥▼P ❈✐❧❦ ❈❯❉❆ ❖♣❡♥❈▲ ❈❤❛r♠✰✰

❑❆❆P■ ❙t❛rP❯ ◗❯❆❘❑ P❛r❙❊❈ ❖♠♣❙s ✳✳✳

♥❡❡❞ ❢♦r ❛ ♣♦rt❛❜❧❡ ❛♥❞ ❡✣❝✐❡♥t ✇❛② t♦ ❡①♣❧♦✐t s✉❝❤ ❛r❝❤✐t❡❝t✉r❡s

✶✹ ✴ ✸✵

❙t❛rP❯ ❛♥❞ ❙✐♠●r✐❞

❙t❛rP❯ ✭■♥r✐❛ ❇♦r❞❡❛✉①✮

Dynamic runtime for hybrid architectures (CPU, GPU, MPI)

Opportunistic scheduling of a task graph guided by resource performancemodels

Features both dense and sparse applications. FMM ongoing.

❙✐♠●r✐❞

Scalable Simulation framework for distributed systems

Sound fluid network models accounting for heterogeneity and contention

Modeling with threads rather than only trace replay ability to simulatedynamic applications

Portable, open source and easily extendable

❙t❛rP❯ ✇❛s ♣♦rt❡❞ ♦♥ t♦♣ ♦❢ ❙✐♠●r✐❞ ❜② ❙✳ ❚❤✐❜❛✉❧t ✐♥ ✶ ❞❛②✿Replace synchronization and thread creation by SimGrid’s onesVery crude platform model

❚❤❡ s❛♠❡ ❛♣♣r♦❛❝❤ s❤♦✉❧❞ ❜❡ ❛♣♣❧✐❝❛❜❧❡ t♦ ❛♥② t❛s❦✲❜❛s❡❞ r✉♥t✐♠❡✶✺ ✴ ✸✵

❊♥✈✐s✐♦♥❡❞ ❲♦r❦✢♦✇✿ ❙t❛rP❯✰❙✐♠●r✐❞

❙t❛rP❯

❙✐♠●r✐❞

❙✐♠✉❧❛t✐♦♥

◗✉✐❝❦❧② ❙✐♠✉❧❛t❡ ▼❛♥② ❚✐♠❡s

❙t❛rP❯

P❡r❢♦r♠❛♥❝❡ Pr♦✜❧❡

❈❛❧✐❜r❛t✐♦♥

❘✉♥ ♦♥❝❡✦✶✻ ✴ ✸✵

❊♥✈✐s✐♦♥❡❞ ❲♦r❦✢♦✇✿ ❙t❛rP❯✰❙✐♠●r✐❞

❙t❛rP❯

❙✐♠●r✐❞

❙✐♠✉❧❛t✐♦♥

◗✉✐❝❦❧② ❙✐♠✉❧❛t❡ ▼❛♥② ❚✐♠❡s

❙t❛rP❯

P❡r❢♦r♠❛♥❝❡ Pr♦✜❧❡

❈❛❧✐❜r❛t✐♦♥

❘✉♥ ♦♥❝❡✦✶✻ ✴ ✸✵

■♠♣❧❡♠❡♥t❛t✐♦♥ Pr✐♥❝✐♣❧❡s

❊♠✉❧❛t✐♦♥ ❡①❡❝✉t✐♥❣ r❡❛❧ ❛♣♣❧✐❝❛t✐♦♥s ✐♥ ❛ s②♥t❤❡t✐❝ ❡♥✈✐r♦♥♠❡♥t✱ ❣❡♥❡r❛❧❧②s❧♦✇✐♥❣ ❞♦✇♥ t❤❡ ✇❤♦❧❡ ❝♦❞❡

❙✐♠✉❧❛t✐♦♥ ✉s❡ ❛ ♣❡r❢♦r♠❛♥❝❡ ♠♦❞❡❧ t♦ ❞❡t❡r♠✐♥❡ ❤♦✇ ♠✉❝❤ t✐♠❡ ❛♣r♦❝❡ss s❤♦✉❧❞ ✇❛✐t

❙t❛rP❯ ❛♣♣❧✐❝❛t✐♦♥s ❛♥❞ r✉♥t✐♠❡ ❛r❡ ❡♠✉❧❛t❡❞ ✭r❡❛❧ s❝❤❡❞✉❧❡r ❛♥❞❞②♥❛♠✐❝ ❞❡❝✐s✐♦♥ ❣✉✐❞❡❞ ♦♥ ❙t❛rP❯ ❝❛❧✐❜r❛t✐♦♥✮

❆❧❧ ♦♣❡r❛t✐♦♥s r❡❧❛t❡❞ t♦ t❤r❡❛❞ s②♥❝❤r♦♥✐③❛t✐♦♥✱ ❛❝t✉❛❧ ❝♦♠♣✉t❛t✐♦♥s✱♠❡♠♦r② ❛❧❧♦❝❛t✐♦♥ ❛♥❞ ❞❛t❛ tr❛♥s❢❡r ❛r❡ s✐♠✉❧❛t❡❞ ✭♥❡❡❞ ❢♦r ❛ ❣♦♦❞❦❡r♥❡❧ ❛♥❞ ❝♦♠♠✉♥✐❝❛t✐♦♥ ♠♦❞❡❧✮ ❛♥❞ ❢❛❦❡❞

Actual computation results are irrelevant and have no impact on thecontrol flow. Only time mattersIn SimGrid, all threads run in mutual exclusion (polling)

❚❤❡ ❝♦♥tr♦❧ ♣❛rt ♦❢ ❙t❛rP❯ ✐s ♠♦❞✐✜❡❞ t♦ ❞②♥❛♠✐❝❛❧❧② ✐♥❥❡❝t❝♦♠♣✉t❛t✐♦♥ ❛♥❞ ❝♦♠♠✉♥✐❝❛t✐♦♥ t❛s❦s ✐♥t♦ t❤❡ s✐♠✉❧❛t♦r

✶✼ ✴ ✸✵

❉❡♥s❡ ▲✐♥❡❛r ❆❧❣❡❜r❛ ❆♣♣❧✐❝❛t✐♦♥s

❙t❛rt❡❞ ✇✐t❤ r❡❣✉❧❛r ❞❡♥s❡ ❦❡r♥❡❧s ❛♥❞ ❛ ✜①❡❞ t✐❧❡ s✐③❡❯s❡❞ t✇♦ ❞✐✛❡r❡♥t ♠❛tr✐① ❞❡❝♦♠♣♦s✐t✐♦♥ ❛❧❣♦r✐t❤♠s✿

✶ Cholesky✷ LU

❯s❡❞ ❛ ✇✐❞❡ ❞✐✈❡rs✐t② ♦❢ ♠❛❝❤✐♥❡s

◆❛♠❡ Pr♦❝❡ss♦r ★❈♦r❡s ▼❡♠♦r② ●P❯s

❤❛♥♥✐❜❛❧ ❳✺✺✺✵ ✷× ✹ ✷× ✷✹●❇ ✸×◗✉❛❞r♦❋❳✺✽✵✵❛tt✐❧❛ ❳✺✻✺✵ ✷× ✻ ✷× ✷✹●❇ ✸×❚❡s❧❛❈✷✵✺✵♠✐r❛❣❡ ❳✺✻✺✵ ✷× ✻ ✷× ✶✽●❇ ✸×❚❡s❧❛▼✷✵✼✵❝♦♥❛♥ ❊✺✲✷✻✺✵ ✷× ✽ ✷× ✸✷●❇ ✸×❚❡s❧❛▼✷✵✼✺❢r♦❣❦❡♣❧❡r ❊✺✲✷✻✼✵ ✷× ✽ ✷× ✶✻●❇ ✷×❑✷✵♣✐❧✐♣✐❧✐✷ ❊✺✲✷✻✸✵ ✷× ✻ ✷× ✸✷●❇ ✷×❑✹✵✐❞❣r❛❢ ❳✺✻✺✵ ✷× ✻ ✷× ✸✻●❇ ✽×❚❡s❧❛❈✷✵✺✵✐❞❝❤✐r❡ ❊✺✲✹✻✹✵ ✷✹× ✽ ✷✹× ✸✶●❇ ✴

Table: Machines used for the dense linear algebra experiments.

✶✽ ✴ ✸✵

❚❤❡ ♣❛t❤ t♦ r❡❧✐❛❜❧❡ ♣r❡❞✐❝t✐♦♥s

Conan Cholesky Attila LU

0

500

1000

1500

20K 40K 60K 80K 20K 40K 60K 80KMatrix dimension

GF

lop/

s

ExperimentalCondition

SimGrid (improved)

SimGrid (initial)

Native

Hannibal LU QuadroFX5800

0

250

500

750

20K 40K 60K 80KMatrix dimension

GF

lop/

s

ExperimentalCondition

SimGrid (smart)

SimGrid (improved)

SimGrid (initial)

Native

●❡tt✐♥❣ ❡①❝❡❧❧❡♥t r❡s✉❧ts ✭❡✳❣✳✱ ❈❤♦❧❡s❦② ♦♥ ❈♦♥❛♥✮ s♦♠❡t✐♠❡s ❞♦ ♥♦tr❡q✉✐r❡s ♠✉❝❤ ❡✛♦rts

❇✉t ♠♦❞❡❧✐♥❣ ❝♦♠♠✉♥✐❝❛t✐♦♥ ❤❡t❡r♦❣❡♥❡✐t②✱ ❝♦♥t❡♥t✐♦♥✱ ♠❡♠♦r②♦♣❡r❛t✐♦♥ ✭❛♥❞ ❡✈❡♥ s♦♠❡t✐♠❡s ❤❛r❞✇❛r❡✴❞r✐✈❡r ♣❡❝✉❧✐❛r✐t②✮ ✐s ❡ss❡♥t✐❛❧

❚r② t♦ ❜❡ ❛s ❡①❤❛✉st✐✈❡ ❛s ♣♦ss✐❜❧❡✳ ✳ ✳

✶✾ ✴ ✸✵

❖✈❡r✈✐❡✇ ♦❢ ❙✐♠✉❧❛t✐♦♥ ❆❝❝✉r❛❝②

hannibal: 3 QuadroFX5800 attila: 3 TeslaC2050 mirage: 3 TeslaM2070

0

1000

2000

3000

4000

0

1000

2000

3000

4000

Ch

ole

sky

LU

20K 40K 60K 80K 20K 40K 60K 80K 20K 40K 60K 80K

Matrix dimension

GF

lop

/s

ExperimentalCondition

SimGrid

Native

Checking predictive capability of the simulation

conan: 3 TeslaM2075 frogkepler: 2 K20 pilipili2: 2 K40

0

1000

2000

3000

4000

0

1000

2000

3000

4000

Ch

ole

sky

LU

20K 40K 60K 80K 20K 40K 60K 80K 20K 40K 60K 80K

Matrix dimension

GF

lop

/s

ExperimentalCondition

SimGrid

Native

✷✵ ✴ ✸✵

❇❡②♦♥❞ ❙✐♠♣❧❡ ●r❛♣❤s

❈♦♠♣❛r✐♥❣ ❉✐✛❡r❡♥t ❙❝❤❡❞✉❧❡rs

❈❤♦❧❡s❦② ♦♥ ❆tt✐❧❛

DMDA DMDAR DMDAS

0

500

1000

1500

20K 40K 60K 80K 20K 40K 60K 80K 20K 40K 60K 80K

Matrix dimension

GF

lop

/s

ExperimentalCondition

SimGrid

Native

■♥✈❡st✐❣❛t✐♥❣ ❉❡t❛✐❧s

✷✶ ✴ ✸✵

❖✉t❧✐♥❡

✶ ❈♦♥t❡①t

✷ ❙✐♠●r✐❞ ❛♥❞ ▼P■✿ ❙▼P■❙▼P■❙✐♠●r✐❞ ❛♥❞ ❆▼P■

✸ ❙✐♠●r✐❞ ❛♥❞ ❙t❛rP❯Pr✐♥❝✐♣❧❡❉❡♥s❡ ▲✐♥❡❛r ❆❧❣❡❜r❛ ❆♣♣❧✐❝❛t✐♦♥s❙♣❛rs❡ ▲✐♥❡❛r ❆❧❣❡❜r❛ ❆♣♣❧✐❝❛t✐♦♥s

✹ ❈♦♥❝❧✉s✐♦♥

❙✐♠✉❧❛t✐♥❣ ❙♣❛rs❡ ❙♦❧✈❡rs

qr♠❴st❛r♣✉

◗❘ ▼❯▼P❙ ♠✉❧t✐✲❢r♦♥t❛❧ ❢❛❝t♦r✐③❛t✐♦♥ ♦♥ t♦♣ ♦❢ ❙t❛rP❯

Tree parallelism: nodes inseparate branches can be treatedindependently

Node parallelism: large nodescan be treated by multipleprocess

Activate

Assemble

Panel

UpdateDeactivate

◆♦ ●P❯ s✉♣♣♦rt ✭♦♥❣♦✐♥❣✮ ✐♥ t❤✐s st✉❞②✱ ♦♥❧② ♠✉❧t✐✲❝♦r❡

P♦rt✐♥❣ qr♠❴st❛r♣✉ ♦♥ t♦♣ ♦❢ ❙✐♠●r✐❞

❈❤❛♥❣✐♥❣ ♠❛✐♥ ❢♦r t❤❡ s✉❜r♦✉t✐♥❡

❈❤❛♥❣✐♥❣ ❝♦♠♣✐❧❛t✐♦♥ ♣r♦❝❡ss

❈❛r❡❢✉❧ ❦❡r♥❡❧ ♠♦❞❡❧✐♥❣ ❛s ♠❛tr✐① ❞✐♠❡♥s✐♦♥ ❦❡❡♣s ❝❤❛♥❣✐♥❣

✷✷ ✴ ✸✵

❊①❛♠♣❧❡ ❢♦r ▼♦❞❡❧✐♥❣ ❑❡r♥❡❧s✿ ●❊◗❘❚

●❊◗❘❚✭P❛♥❡❧✮ ❞✉r❛t✐♦♥✿

T●❊◗❘❚ = a+ ✷b(◆❇✷×▼❇)− ✷c(◆❇✸

× ❇❑ ) +✹d

✸◆❇

✸

❲❡ ❝❛♥ ❞♦ ❛ ❧✐♥❡❛r r❡❣r❡ss✐♦♥ ❜❛s❡❞ ♦♥ ❛❞ ❤♦❝ ❝❛❧✐❜r❛t✐♦♥

●❊◗❘❚ ❉✉r❛t✐♦♥

NB✸ ✶.✺✵× ✶✵−✺ ✭✶.✸✵× ✶✵−✺✱ ✶.✼✵× ✶✵−✺✮ ∗∗∗

NB✷∗MB ✺.✹✾× ✶✵−✼ ✭✺.✹✻× ✶✵−✼✱ ✺.✺✶× ✶✵−✼✮ ∗∗∗

NB✸∗ BK −✺.✺✷× ✶✵−✼ ✭−✺.✺✼× ✶✵−✼✱ −✺.✹✽× ✶✵−✼✮ ∗∗∗

❈♦♥st❛♥t −✷.✹✾× ✶✵✶ ✭−✷.✽✸× ✶✵✶✱ −✷.✶✹× ✶✵✶✮ ∗∗∗

❖❜s❡r✈❛t✐♦♥s ✹✾✸R✷ ✵.✾✾✾

◆♦t❡✿ ∗♣<✵✳✶❀ ∗∗♣<✵✳✵✺❀ ∗∗∗♣<✵✳✵✶

✷✸ ✴ ✸✵

❈♦♠♣❛r✐♥❣ ❑❡r♥❡❧ ❉✉r❛t✐♦♥ ❉✐str✐❜✉t✐♦♥s

❉♦❴s✉❜tr❡❡ ■◆■❚ ●❊◗❘❚ ●❊▼◗❘❚ ❆❙▼

✶✳ ★❋❧♦♣s ★❩❡r♦s ◆❇ ◆❇ ★❈♦❡✛✷✳ ★◆♦❞❡s ★❆ss❡♠❜❧❡ ▼❇ ▼❇ ✴✸✳ ✴ ✴ ❇❑ ❇❑ ✴

R✷ ✵✳✾✾ ✵✳✾✾ ✵✳✾✾ ✵✳✾✾ ✵✳✽✻

Native, Do_subtree Native, INIT Native, GEQRT Native, GEMQRT Native, ASM Native, CLEAN

SimGrid, Do_subtree SimGrid, INIT SimGrid, GEQRT SimGrid, GEMQRT SimGrid, ASM SimGrid, CLEAN

0

2

4

6

0

5

10

15

20

0

50

100

150

200

0

2000

4000

0

50

100

150

200

250

0

10

20

0

2

4

6

0

5

10

15

0

100

200

0

2000

4000

6000

0

100

200

300

0

10

20

30

0 50 100 0 250 500 750 0 25 50 75 100 0 20 40 60 0 10 20 30 0 25 50 75

50 100 0 250 500 750 0 25 50 75 0 20 40 60 0 10 20 0 25 50 75Kernel Duration [ms]

Num

ber

of O

ccur

ance

s Kernel

Do_subtree

INIT

GEQRT

GEMQRT

ASM

CLEAN

✷✹ ✴ ✸✵

❖✈❡r✈✐❡✇ ♦❢ ❙✐♠✉❧❛t✐♦♥ ❆❝❝✉r❛❝②

0

10

20

30

40

50

0

200

400

tp−6

karte

dEt

erni

tyII_

Ede

gme

hirla

me1

8TF

16

Ruc

ci1 sls

TF17

Mak

espa

n [s

]

Type

Native

SimGrid

Fourmi machine with 8 cores

0

10

20

30

40

0

100

200

300

400

500

tp−6

karte

dEt

erni

tyII_

Ede

gme

hirla

me1

8TF

16

Ruc

ci1 sls

TF17

Mak

espa

n [s

]

Type

Native

SimGrid

Riri machine with 10 cores

❘❡s✉❧ts ✐♥ ❛ ♥✉ts❤❡❧❧

▼♦st ♦❢ t❤❡ t✐♠❡✱ s✐♠✉❧❛t✐♦♥ ✐ss❧✐❣❤t❧② ♦♣t✐♠✐st✐❝

❲✐t❤ ❜✐❣❣❡r ❛♥❞ ❛r❝❤✐t❡❝t✉r❛❧❧②♠♦r❡ ❝♦♠♣❧❡① ♠❛❝❤✐♥❡s✱ ❡rr♦r✐♥❝r❡❛s❡s

0

5

10

15

20

0

50

100

150

200

tp−6

karte

dEt

erni

tyII_

Ede

gme

hirla

me1

8TF

16

Ruc

ci1 sls

TF17

Mak

espa

n [s

]

Type

Native

SimGrid

Riri machine with 40 cores

✷✺ ✴ ✸✵

❱✐s✉❛❧ ❚r❛❝❡ ❈♦♠♣❛r✐s♦♥

❚r❛❝❡s ❡①tr❡♠❡❧② ✧❝❧♦s❡✧

❙✐♠✉❧❛t❡❞ ♠❛❦❡s♣❛♥ s❧✐❣❤t❧② ♦♣t✐♠✐st✐❝

❉✐✛❡r❡♥t s❝❤❡❞✉❧✐♥❣ ❛t t❤❡ ❜❡❣✐♥♥✐♥❣ ❛♥❞ ❛t t❤❡ ❡♥❞

✷✻ ✴ ✸✵

❙t✉❞②✐♥❣ ▼❡♠♦r② ❈♦♥s✉♠♣t✐♦♥

▼✐♥✐♠✐③✐♥❣ ♠❡♠♦r② ❢♦♦t♣r✐♥t ✐s ✈❡r② ✐♠♣♦rt❛♥t ❢♦r s✉❝❤ ❛♣♣❧✐❝❛t✐♦♥s❘❡♠❡♠❜❡r s❝❤❡❞✉❧✐♥❣ ✐s ❞②♥❛♠✐❝ s♦ ❝♦♥s❡❝✉t✐✈❡ ◆❛t✐✈❡ ❡①♣❡r✐♠❡♥ts❤❛✈❡ ❞✐✛❡r❡♥t ♦✉t♣✉t

Experiment number 1

Experiment number 2

Experiment number 3

Experiment number 4

0

1

2

3

0

1

2

3

0

1

2

3

0

1

2

3

0 10,000 20,000 30,000 40,000Time [ms]

Allo

cate

d M

emor

y [G

iB]

✷✼ ✴ ✸✵

❙t✉❞②✐♥❣ ▼❡♠♦r② ❈♦♥s✉♠♣t✐♦♥

▼✐♥✐♠✐③✐♥❣ ♠❡♠♦r② ❢♦♦t♣r✐♥t ✐s ✈❡r② ✐♠♣♦rt❛♥t ❢♦r s✉❝❤ ❛♣♣❧✐❝❛t✐♦♥s❘❡♠❡♠❜❡r s❝❤❡❞✉❧✐♥❣ ✐s ❞②♥❛♠✐❝ s♦ ❝♦♥s❡❝✉t✐✈❡ ◆❛t✐✈❡ ❡①♣❡r✐♠❡♥ts❤❛✈❡ ❞✐✛❡r❡♥t ♦✉t♣✉t

Native 1

SimGrid

Native 2

Native 3

0

1

2

3

0

1

2

3

0

1

2

3

0

1

2

3

0 10,000 20,000 30,000 40,000Time [ms]

Allo

cate

d M

emor

y [G

iB]

✷✼ ✴ ✸✵

❊①tr❛♣♦❧❛t✐♥❣ t♦ ▲❛r❣❡r ▼❛❝❤✐♥❡s

Pr❡❞✐❝t✐♥❣ ♣❡r❢♦r♠❛♥❝❡✐♥ ✐❞❡❛❧✐③❡❞ ❝♦♥t❡①t

❙t✉❞②✐♥❣ t❤❡♣❛r❛❧❧❡❧✐③❛t✐♦♥ ❧✐♠✐ts ♦❢t❤❡ ♣r♦❜❧❡♠

Extrapolation

0

30

60

90

4 10 20 40 100 400Number of Threads

Dur

atio

n [s

]

Type

Native

SimGrid

Measured Time

Overall Makespan

Idle Time per Thread

❘❡❣❛r❞✐♥❣ s✐♠✉❧❛t✐♦♥ s♣❡❡❞✿

◗✉✐❝❦❧② ❛♥❞ ❛❝❝✉r❛t❡❧② ❡✈❛❧✉❛t❡ t❤❡ ✐♠♣❛❝t ♦❢ ✈❛r✐♦✉ss❝❤❡❞✉❧✐♥❣✴❛♣♣❧✐❝❛t✐♦♥ ♣❛r❛♠❡t❡rs✿

qr♠❴st❛r♣✉ ❈♦r❡s ❘❆▼ ❊✈❛❧✉❛t✐♦♥ ▼❛❦❡s♣❛♥

◆❛t✐✈❡ ✹✵ ✺✽✳✵●✐❇ ✶✺✼s ✶✹✶s❙✐♠●r✐❞ ✶ ✶✳✺●✐❇ ✺✼s ✶✹✷s

✷✽ ✴ ✸✵

❈♦♥❝❧✉s✐♦♥

❲❡ ❤❛✈❡ ♥♦✇ ❛❝❝✉r❛t❡ ❜❛s❡❧✐♥❡s t♦ ❝♦♠♣❛r❡ ✇✐t❤ ✇❤❡♥❡✈❡r t❤❡r❡ ✐s❛ ♠✐s♠❛t❝❤✱ ✇❡ ❝❛♥ q✉❡st✐♦♥ s✐♠✉❧❛t✐♦♥ ❛s ✇❡❧❧ ❛s ❡①♣❡r✐♠❡♥t❛❧ s❡t✉♣✿

TCP RTO issueInaccurate platform specifications

Flawed MPI optimization

❍♦♣❡ ✐t ✇✐❧❧ ❜❡ ✉s❡❢✉❧ t♦the Mont-Blanc projectyou?. . .

BigDFT or Ondes3D developers

◆❡❡❞ t♦ ✈❛❧✐❞❛t❡ t❤✐s ❛♣♣r♦❛❝❤ ♦♥ ❧❛r❣❡r ♣❧❛t❢♦r♠s✱ ✇✐t❤ ♦t❤❡r ♥❡t✇♦r❦t②♣❡s ❛♥❞ t♦♣♦❧♦❣✐❡s ✭❡✳❣✳✱ ■♥✜♥✐❜❛♥❞✱ t♦r✉s✮❈♦♠♠✉♥✐❝❛t✐♦♥ t❤r♦✉❣❤ s❤❛r❡❞ ♠❡♠♦r② ✐s ♦❦✱ ❜✉t ♠♦❞❡❧✐♥❣ t❤❡✐♥t❡r❢❡r❡♥❝❡ ❜❡t✇❡❡♥ ♠❡♠♦r②✲❜♦✉♥❞ ❦❡r♥❡❧s ✐s r❡❛❧❧② ❤❛r❞❙▼P■ ❛♥❞ ❙t❛rP❯✴❙✐♠●r✐❞ ❛r❡ ♦♣❡♥ s♦✉r❝❡✴s❝✐❡♥❝❡✿ ✇❡ ♣✉t ❛ ❧♦t ♦❢❡✛♦rt ✐♥t♦ ♠❛❦✐♥❣ ✐t ✉s❛❜❧❡

❤tt♣✿✴✴s✐♠❣r✐❞✳❣❢♦r❣❡✳✐♥r✐❛✳❢r

✷✾ ✴ ✸✵

❖♥❣♦✐♥❣ ❲♦r❦

⊠ ❙❡❛♠❧❡ss ❡♠✉❧❛t✐♦♥ ✭♠♠❛♣ ❛♣♣r♦❛❝❤ ✇♦r❦s ❣r❡❛t✮

⊟ ▼♦❞❡❧✐♥❣ ■❇ ♥❡t✇♦r❦s✱ t♦r✉s✴❢❛t tr❡❡ t♦♣♦❧♦❣✐❡s

⊟ ▼♦❞❡❧✐♥❣ ❡♥❡r❣② ✭✇✐t❤ ❆✳❈✳ ❖r❣❡r✐❡✮

⊠ ❘✉♥t✐♠❡s ❢♦r ❤②❜r✐❞ ✭❈P❯✰●P❯✮ ♣❧❛t❢♦r♠s ✭❙t❛rP❯✱ ✇✐t❤ ❙✳❚❤✐❜❛✉❧t✮ ♦♥ t♦♣ ♦❢ ❙✐♠●r✐❞

Works great for both dense (MAGMA/MORSE) and sparse(QR-MUMPS) linear algebra. Ongoing work for FMM.Large scale hybrid platforms (StarPU-MPI)

⊟ ❋♦r♠❛❧ ✈❡r✐✜❝❛t✐♦♥ ❜② ❡①❤❛✉st✐✈❡ st❛t❡ s♣❛❝❡ ❡①♣❧♦r❛t✐♦♥

⊟ ❈♦♠♣✉t❛t✐♦♥✴❝♦♠♠✉♥✐❝❛t✐♦♥✴♠❡♠♦r② ❛❝❝❡ss ✐♥t❡r❢❡r❡♥❝❡s

� ❖♣❡♥▼P✴♣t❤r❡❛❞s

⊟ ■❖s

✸✵ ✴ ✸✵