optimizing the steady state throughput of scatter and
play

Optimizing the steady-state throughput of scatter and reduce - PowerPoint PPT Presentation

May 14, 2023 •470 likes •1.62k views

Optimizing the steady-state throughput of scatter and reduce operations on heterogeneous platforms Arnaud Legrand, Loris Marchal, Yves Robert Laboratoire de lInformatique du Parall elisme Ecole Normale Sup erieure de Lyon, France

  1. Two Problems of Collective Communication Scatter one processor P source sends distinct messages to target � � processors ( ) P t 0 , . . . , P t N ◮ Series of Scatter P source sends consecutively a large number of distinct messages to all targets Reduce Each of the participating processor P r i in P r 0 , . . . , P r N owns a value v i ⇒ compute V = v 1 ⊕ v 2 ⊕ · · · ⊕ v N ( ⊕ is associative, non commutative) ◮ Series of Reduce several consecutive reduce operations from the same set P r 0 , . . . , P r N to the same target P target . Loris Marchal Steady state collective communications 5/ 27

  2. Two Problems of Collective Communication Scatter one processor P source sends distinct messages to target � � processors ( ) P t 0 , . . . , P t N ◮ Series of Scatter P source sends consecutively a large number of distinct messages to all targets Reduce Each of the participating processor P r i in P r 0 , . . . , P r N owns a value v i ⇒ compute V = v 1 ⊕ v 2 ⊕ · · · ⊕ v N ( ⊕ is associative, non commutative) ◮ Series of Reduce several consecutive reduce operations from the same set P r 0 , . . . , P r N to the same target P target . Loris Marchal Steady state collective communications 5/ 27

  3. Platform Model ◮ G = ( V, E, c ) ◮ P 1 , P 2 , . . . , P n : processors P 1 ◮ ( j, k ) ∈ E : communication link 10 between P i and P j 10 30 ◮ c ( j, k ) : time to transfer one unit P 0 P 3 message from P j to P k 5 5 8 ◮ one-port for incoming P 2 communications ◮ one-port for outgoing communications Loris Marchal Steady state collective communications 6/ 27

  4. Platform Model ◮ G = ( V, E, c ) ◮ P 1 , P 2 , . . . , P n : processors P 1 ◮ ( j, k ) ∈ E : communication link 10 between P i and P j 10 30 ◮ c ( j, k ) : time to transfer one unit P 0 P 3 message from P j to P k 5 5 8 ◮ one-port for incoming P 2 communications ◮ one-port for outgoing communications Loris Marchal Steady state collective communications 6/ 27

  5. Platform Model ◮ G = ( V, E, c ) ◮ P 1 , P 2 , . . . , P n : processors P 1 ◮ ( j, k ) ∈ E : communication link 10 between P i and P j 10 30 ◮ c ( j, k ) : time to transfer one unit P 0 P 3 message from P j to P k 5 5 8 ◮ one-port for incoming P 2 communications ◮ one-port for outgoing communications Loris Marchal Steady state collective communications 6/ 27

  6. Platform Model ◮ G = ( V, E, c ) ◮ P 1 , P 2 , . . . , P n : processors P 1 ◮ ( j, k ) ∈ E : communication link 10 between P i and P j 10 30 ◮ c ( j, k ) : time to transfer one unit P 0 P 3 message from P j to P k 5 5 8 ◮ one-port for incoming P 2 communications ◮ one-port for outgoing communications Loris Marchal Steady state collective communications 6/ 27

  7. Platform Model ◮ G = ( V, E, c ) ◮ P 1 , P 2 , . . . , P n : processors P 1 ◮ ( j, k ) ∈ E : communication link 10 between P i and P j 10 30 ◮ c ( j, k ) : time to transfer one unit P 0 P 3 message from P j to P k 5 5 8 ◮ one-port for incoming P 2 communications ◮ one-port for outgoing communications Loris Marchal Steady state collective communications 6/ 27

  8. Platform Model ◮ G = ( V, E, c ) ◮ P 1 , P 2 , . . . , P n : processors P 1 ◮ ( j, k ) ∈ E : communication link 10 between P i and P j 10 30 ◮ c ( j, k ) : time to transfer one unit P 0 P 3 message from P j to P k 5 5 8 ◮ one-port for incoming P 2 communications ◮ one-port for outgoing communications Loris Marchal Steady state collective communications 6/ 27

  9. Platform Model ◮ G = ( V, E, c ) ◮ P 1 , P 2 , . . . , P n : processors P 1 ◮ ( j, k ) ∈ E : communication link between P i and P j ◮ c ( j, k ) : time to transfer one unit P 0 P 3 message from P j to P k ◮ one-port for incoming P 2 communications ◮ one-port for outgoing communications Loris Marchal Steady state collective communications 6/ 27

  10. Platform Model ◮ G = ( V, E, c ) ◮ P 1 , P 2 , . . . , P n : processors P 1 ◮ ( j, k ) ∈ E : communication link OK between P i and P j ◮ c ( j, k ) : time to transfer one unit P 0 P 3 message from P j to P k ◮ one-port for incoming P 2 communications ◮ one-port for outgoing communications Loris Marchal Steady state collective communications 6/ 27

  11. Framework 1. express optimization problem as set of linear constraints (variables = fraction of time a processor spends sending to one of its neighbors) 2. solve linear program (in rational numbers) 3. use solution to build periodic schedule reaching best throughput Loris Marchal Steady state collective communications 7/ 27

  12. Framework 1. express optimization problem as set of linear constraints (variables = fraction of time a processor spends sending to one of its neighbors) 2. solve linear program (in rational numbers) 3. use solution to build periodic schedule reaching best throughput Loris Marchal Steady state collective communications 7/ 27

  13. Framework 1. express optimization problem as set of linear constraints (variables = fraction of time a processor spends sending to one of its neighbors) 2. solve linear program (in rational numbers) 3. use solution to build periodic schedule reaching best throughput Loris Marchal Steady state collective communications 7/ 27

  14. Outline Introduction Two Problems of Collective Communication Platform Model Framework Series of Scatter Steady-state constraints Toy Example Building a schedule Asymptotic optimality Series of Reduce Introduction to reduction trees Linear Program Periodic schedule - Asymptotic optimality Toy Example for Series of Reduce Approximation for a fixed period Conclusion Loris Marchal Steady state collective communications 8/ 27

  15. Series of Scatter ◮ m k : types of the messages with destination P k ◮ s ( P i → P j , m k ) : fractional number of messages of type m k sent on the edge P i → P j within on time unit ◮ t ( P i → P j ) : fractional time spent by processor P i to send data to its neighbor P j within one time unit ◮ bound for this activity: ∀ P i , P j , 0 � t ( P i → P j ) � 1 ◮ on a link P i → P j during one time-unit: � t ( P i → P j ) = s ( P i → P j , m k ) k Loris Marchal Steady state collective communications 9/ 27

  16. Series of Scatter ◮ m k : types of the messages with destination P k ◮ s ( P i → P j , m k ) : fractional number of messages of type m k sent on the edge P i → P j within on time unit ◮ t ( P i → P j ) : fractional time spent by processor P i to send data to its neighbor P j within one time unit ◮ bound for this activity: ∀ P i , P j , 0 � t ( P i → P j ) � 1 ◮ on a link P i → P j during one time-unit: � t ( P i → P j ) = s ( P i → P j , m k ) k Loris Marchal Steady state collective communications 9/ 27

  17. Series of Scatter ◮ m k : types of the messages with destination P k ◮ s ( P i → P j , m k ) : fractional number of messages of type m k sent on the edge P i → P j within on time unit ◮ t ( P i → P j ) : fractional time spent by processor P i to send data to its neighbor P j within one time unit ◮ bound for this activity: ∀ P i , P j , 0 � t ( P i → P j ) � 1 ◮ on a link P i → P j during one time-unit: � t ( P i → P j ) = s ( P i → P j , m k ) k Loris Marchal Steady state collective communications 9/ 27

  18. Series of Scatter ◮ m k : types of the messages with destination P k ◮ s ( P i → P j , m k ) : fractional number of messages of type m k sent on the edge P i → P j within on time unit ◮ t ( P i → P j ) : fractional time spent by processor P i to send data to its neighbor P j within one time unit ◮ bound for this activity: ∀ P i , P j , 0 � t ( P i → P j ) � 1 ◮ on a link P i → P j during one time-unit: � t ( P i → P j ) = s ( P i → P j , m k ) k Loris Marchal Steady state collective communications 9/ 27

  19. Series of Scatter ◮ m k : types of the messages with destination P k ◮ s ( P i → P j , m k ) : fractional number of messages of type m k sent on the edge P i → P j within on time unit ◮ t ( P i → P j ) : fractional time spent by processor P i to send data to its neighbor P j within one time unit ◮ bound for this activity: ∀ P i , P j , 0 � t ( P i → P j ) � 1 ◮ on a link P i → P j during one time-unit: � t ( P i → P j ) = s ( P i → P j , m k ) k Loris Marchal Steady state collective communications 9/ 27

  20. Linear constraints ◮ one port constraints for outgoing messages in P i : � ∀ P i , t ( P i → P j ) � 1 P i → P j ◮ one port constraints for incoming messages in P i : � ∀ P i , t ( P j → P i ) � 1 P j → P i ◮ conservation law in node P i for message m k ( k � = i ): 5 m k 3 m k P i 4 m k 2 m k � � s ( P j → P i , m k ) = s ( P j → P i , m k ) � 1 P j → P i P i → P j Loris Marchal Steady state collective communications 10/ 27

  21. Linear constraints ◮ one port constraints for outgoing messages in P i : � ∀ P i , t ( P i → P j ) � 1 P i → P j ◮ one port constraints for incoming messages in P i : � ∀ P i , t ( P j → P i ) � 1 P j → P i ◮ conservation law in node P i for message m k ( k � = i ): 5 m k 3 m k P i 4 m k 2 m k � � s ( P j → P i , m k ) = s ( P j → P i , m k ) � 1 P j → P i P i → P j Loris Marchal Steady state collective communications 10/ 27

  22. Linear constraints ◮ one port constraints for outgoing messages in P i : � ∀ P i , t ( P i → P j ) � 1 P i → P j ◮ one port constraints for incoming messages in P i : � ∀ P i , t ( P j → P i ) � 1 P j → P i ◮ conservation law in node P i for message m k ( k � = i ): 5 m k 3 m k P i 4 m k 2 m k � � s ( P j → P i , m k ) = s ( P j → P i , m k ) � 1 P j → P i P i → P j Loris Marchal Steady state collective communications 10/ 27

  23. Linear constraints ◮ one port constraints for outgoing messages in P i : � ∀ P i , t ( P i → P j ) � 1 P i → P j ◮ one port constraints for incoming messages in P i : � ∀ P i , t ( P j → P i ) � 1 P j → P i ◮ conservation law in node P i for message m k ( k � = i ): 5 m k 3 m k P i 4 m k 2 m k � � s ( P j → P i , m k ) = s ( P j → P i , m k ) � 1 P j → P i P i → P j Loris Marchal Steady state collective communications 10/ 27

  24. Linear constraints ◮ one port constraints for outgoing messages in P i : � ∀ P i , t ( P i → P j ) � 1 P i → P j ◮ one port constraints for incoming messages in P i : � ∀ P i , t ( P j → P i ) � 1 P j → P i ◮ conservation law in node P i for message m k ( k � = i ): 5 m k 3 m k P i 4 m k 2 m k � � s ( P j → P i , m k ) = s ( P j → P i , m k ) � 1 P j → P i P i → P j Loris Marchal Steady state collective communications 10/ 27

  25. Linear constraints ◮ one port constraints for outgoing messages in P i : � ∀ P i , t ( P i → P j ) � 1 P i → P j ◮ one port constraints for incoming messages in P i : � ∀ P i , t ( P j → P i ) � 1 P j → P i ◮ conservation law in node P i for message m k ( k � = i ): 5 m k 3 m k P i 4 m k 2 m k � � s ( P j → P i , m k ) = s ( P j → P i , m k ) � 1 P j → P i P i → P j Loris Marchal Steady state collective communications 10/ 27

  26. Throughput and Linear Program ◮ throughput: total number of messages m k received in P k � TP = s ( P j → P k , m k ) P j → P k (same throughput for every target node P k ) ◮ summarize this constraints in a linear program: Steady-State Scatter Problem on a Graph SSSP(G) Maximize TP , subject to  ∀ P i , ∀ P j , 0 � s ( P i → P j ) � 1  ∀ P i , �  P j , ( i,j ) ∈ E s ( P i → P j ) � 1     ∀ P i , � P j , ( j,i ) ∈ E s ( P j → P i ) � 1     ∀ P i , P j , s ( P i → P j ) = � m k send ( P i → P j , m k ) × c ( i, j ) ∀ P i , ∀ m k , k � = i, � P j , ( j,i ) ∈ E send ( P j → P i , m k )      = � P j , ( i,j ) ∈ E send ( P i → P j , m k )     ∀ P k , k ∈ T � P i , ( i,k ) ∈ E send ( P i → P k , m k ) = TP  Loris Marchal Steady state collective communications 11/ 27

  27. Throughput and Linear Program ◮ throughput: total number of messages m k received in P k � TP = s ( P j → P k , m k ) P j → P k (same throughput for every target node P k ) ◮ summarize this constraints in a linear program: Steady-State Scatter Problem on a Graph SSSP(G) Maximize TP , subject to  ∀ P i , ∀ P j , 0 � s ( P i → P j ) � 1  ∀ P i , �  P j , ( i,j ) ∈ E s ( P i → P j ) � 1     ∀ P i , � P j , ( j,i ) ∈ E s ( P j → P i ) � 1     ∀ P i , P j , s ( P i → P j ) = � m k send ( P i → P j , m k ) × c ( i, j ) ∀ P i , ∀ m k , k � = i, � P j , ( j,i ) ∈ E send ( P j → P i , m k )      = � P j , ( i,j ) ∈ E send ( P i → P j , m k )     ∀ P k , k ∈ T � P i , ( i,k ) ∈ E send ( P i → P k , m k ) = TP  Loris Marchal Steady state collective communications 11/ 27

  28. Series of Scatter - Toy Example P s 1 1 P a P b 2/3 4/3 4/3 P 0 P 1 platform graph (edges labeled with c ( i, j ) ) Loris Marchal Steady state collective communications 12/ 27

  29. Series of Scatter - Toy Example P s 1 4 m 0 1 4 m 0 1 2 m 1 P a P b 1 4 m 0 1 1 4 m 0 2 m 1 P 0 P 1 value of s ( P i → P j , m k ) in the solution of the linear program Loris Marchal Steady state collective communications 12/ 27

  30. Series of Scatter - Toy Example P s 1/4 3/4 P a P b 1/3 1/6 2/3 P 0 P 1 occupation time of the edge ( t ( P i → P j ) ) Loris Marchal Steady state collective communications 12/ 27

  31. Building a schedule P s ◮ consider the time needed 1 2 ( 1 2 m 1 ) 4 ( 1 1 4 m 0 ) for all transfers 1 4 ( 1 4 m 0 ) ◮ build a bipartite graph by P a P b splitting all nodes 0 ) 4 m 1 6 ( 1 3 ( 1 2 4 m 0 ) 1 2 m 1 ) ◮ extract matchings, using 3 ( 1 the weighted-edge coloring algorithm P 0 P 1 Loris Marchal Steady state collective communications 13/ 27

  32. Building a schedule P send P recv s s ◮ consider the time needed 1 1 1 for all transfers 4 2 4 ◮ build a bipartite graph by P send P recv P send P recv a a b b splitting all nodes 1 ◮ extract matchings, using 1 3 2 6 3 the weighted-edge coloring algorithm P send P recv P send P recv 1 0 1 1 Loris Marchal Steady state collective communications 13/ 27

  33. Building a schedule P send P recv s s ◮ consider the time needed 1 for all transfers 2 ◮ build a bipartite graph by P send P recv P send P recv a a b b splitting all nodes ◮ extract matchings, using 1 2 the weighted-edge coloring algorithm P send P recv P send P recv 1 0 1 1 Loris Marchal Steady state collective communications 13/ 27

  34. Building a schedule P send P recv s s ◮ consider the time needed 1 for all transfers 4 ◮ build a bipartite graph by P send P recv P send P recv a a b b splitting all nodes 1 ◮ extract matchings, using 4 the weighted-edge coloring algorithm P send P recv P send P recv 1 0 1 1 Loris Marchal Steady state collective communications 13/ 27

  35. Building a schedule P send P recv s s ◮ consider the time needed 1 for all transfers 6 ◮ build a bipartite graph by P send P recv P send P recv a a b b splitting all nodes ◮ extract matchings, using 1 1 6 6 the weighted-edge coloring algorithm P send P recv P send P recv 1 0 1 1 Loris Marchal Steady state collective communications 13/ 27

  36. Building a schedule P send P recv s s ◮ consider the time needed 1 for all transfers 12 ◮ build a bipartite graph by P send P recv P send P recv a a b b splitting all nodes 1 ◮ extract matchings, using 12 the weighted-edge coloring algorithm P send P recv P send P recv 1 0 1 1 Loris Marchal Steady state collective communications 13/ 27

  37. Building a schedule matchings: 1 P b → P 1 P send P recv s s 1 P b → P 0 2 P send P recv P send P recv a a b b P a → P 0 1 2 P s → P b P send P recv P send P recv 1 0 1 1 P s → P a t 1 3 11 0 1 2 4 12 ◮ least common multiple T = lcm { b i } where a i b i denotes the number of messages transfered in each matching ◮ ⇒ periodic schedule of period T with atomic transfers of messages Loris Marchal Steady state collective communications 14/ 27

  38. Building a schedule matchings: 2 P b → P 1 P send P recv s s 1 P b → P 0 4 P send P recv P send P recv a a b b P a → P 0 1 4 P s → P b P send P recv P send P recv 1 0 1 1 P s → P a t 1 3 11 0 1 2 4 12 ◮ least common multiple T = lcm { b i } where a i b i denotes the number of messages transfered in each matching ◮ ⇒ periodic schedule of period T with atomic transfers of messages Loris Marchal Steady state collective communications 14/ 27

  39. Building a schedule matchings: 3 P b → P 1 P send P recv s s 1 P b → P 0 6 P send P recv P send P recv a a b b P a → P 0 1 1 6 6 P s → P b P send P recv P send P recv 1 0 1 1 P s → P a t 1 3 11 0 1 2 4 12 ◮ least common multiple T = lcm { b i } where a i b i denotes the number of messages transfered in each matching ◮ ⇒ periodic schedule of period T with atomic transfers of messages Loris Marchal Steady state collective communications 14/ 27

  40. Building a schedule matchings: 4 P b → P 1 P send P recv s s 1 P b → P 0 12 P send P recv P send P recv a a b b P a → P 0 1 12 P s → P b P send P recv P send P recv 1 0 1 1 P s → P a t 1 3 11 0 1 2 4 12 ◮ least common multiple T = lcm { b i } where a i b i denotes the number of messages transfered in each matching ◮ ⇒ periodic schedule of period T with atomic transfers of messages Loris Marchal Steady state collective communications 14/ 27

  41. Building a schedule matchings: 1 2 3 4 P b → P 1 P b → P 0 P a → P 0 P s → P b P s → P a t 1 3 11 0 1 2 4 12 ◮ least common multiple T = lcm { b i } where a i b i denotes the number of messages transfered in each matching ◮ ⇒ periodic schedule of period T with atomic transfers of messages Loris Marchal Steady state collective communications 14/ 27

  42. Building a schedule matchings: 1 2 3 4 P b → P 1 P b → P 0 P a → P 0 P s → P b P s → P a t 1 3 11 0 1 2 4 12 ◮ least common multiple T = lcm { b i } where a i b i denotes the number of messages transfered in each matching ◮ ⇒ periodic schedule of period T with atomic transfers of messages Loris Marchal Steady state collective communications 14/ 27

  43. Building a schedule matchings 1 2 3 4                                                            � P b → P 1 P b → P 0 P a → P 0 P s → P b P s → P a t 0 10 20 30 40 48 ◮ least common multiple T = lcm { b i } where a i b i denotes the number of messages transfered in each matching ◮ ⇒ periodic schedule of period T with atomic transfers of messages Loris Marchal Steady state collective communications 14/ 27

  44. Asymptotic optimality ◮ No schedule can perform more tasks than the steady-state: Lemma. opt ( G, K ) � TP ( G ) × K ◮ periodic schedule ⇒ schedule: 1. initialization phase (fill buffers of messages) 2. r periods of duration T (steady-state) 3. clean-up phase (empty buffers) Lemma. the previous algorithm is asymptotically optimal: steady ( G, K ) lim = 1 opt ( G, K ) K → + ∞ Loris Marchal Steady state collective communications 15/ 27

  45. Asymptotic optimality ◮ No schedule can perform more tasks than the steady-state: Lemma. opt ( G, K ) � TP ( G ) × K ◮ periodic schedule ⇒ schedule: 1. initialization phase (fill buffers of messages) 2. r periods of duration T (steady-state) 3. clean-up phase (empty buffers) Lemma. the previous algorithm is asymptotically optimal: steady ( G, K ) lim = 1 opt ( G, K ) K → + ∞ Loris Marchal Steady state collective communications 15/ 27

  46. Asymptotic optimality ◮ No schedule can perform more tasks than the steady-state: Lemma. opt ( G, K ) � TP ( G ) × K ◮ periodic schedule ⇒ schedule: 1. initialization phase (fill buffers of messages) 2. r periods of duration T (steady-state) 3. clean-up phase (empty buffers) Lemma. the previous algorithm is asymptotically optimal: steady ( G, K ) lim = 1 opt ( G, K ) K → + ∞ Loris Marchal Steady state collective communications 15/ 27

  47. Asymptotic optimality ◮ No schedule can perform more tasks than the steady-state: Lemma. opt ( G, K ) � TP ( G ) × K ◮ periodic schedule ⇒ schedule: 1. initialization phase (fill buffers of messages) 2. r periods of duration T (steady-state) 3. clean-up phase (empty buffers) Lemma. the previous algorithm is asymptotically optimal: steady ( G, K ) lim = 1 opt ( G, K ) K → + ∞ Loris Marchal Steady state collective communications 15/ 27

  48. Asymptotic optimality ◮ No schedule can perform more tasks than the steady-state: Lemma. opt ( G, K ) � TP ( G ) × K ◮ periodic schedule ⇒ schedule: 1. initialization phase (fill buffers of messages) 2. r periods of duration T (steady-state) 3. clean-up phase (empty buffers) Lemma. the previous algorithm is asymptotically optimal: steady ( G, K ) lim = 1 opt ( G, K ) K → + ∞ Loris Marchal Steady state collective communications 15/ 27

  49. Asymptotic optimality ◮ No schedule can perform more tasks than the steady-state: Lemma. opt ( G, K ) � TP ( G ) × K ◮ periodic schedule ⇒ schedule: 1. initialization phase (fill buffers of messages) 2. r periods of duration T (steady-state) 3. clean-up phase (empty buffers) Lemma. the previous algorithm is asymptotically optimal: steady ( G, K ) lim = 1 opt ( G, K ) K → + ∞ Loris Marchal Steady state collective communications 15/ 27

  50. Asymptotic optimality ◮ No schedule can perform more tasks than the steady-state: Lemma. opt ( G, K ) � TP ( G ) × K ◮ periodic schedule ⇒ schedule: 1. initialization phase (fill buffers of messages) 2. r periods of duration T (steady-state) 3. clean-up phase (empty buffers) Lemma. the previous algorithm is asymptotically optimal: steady ( G, K ) lim = 1 opt ( G, K ) K → + ∞ Loris Marchal Steady state collective communications 15/ 27

  51. Outline Introduction Two Problems of Collective Communication Platform Model Framework Series of Scatter Steady-state constraints Toy Example Building a schedule Asymptotic optimality Series of Reduce Introduction to reduction trees Linear Program Periodic schedule - Asymptotic optimality Toy Example for Series of Reduce Approximation for a fixed period Conclusion Loris Marchal Steady state collective communications 16/ 27

  52. Reduce - Reduction trees ◮ Reduce: ◮ each processor P r i owns a value v i ◮ compute V = v 1 ⊕ v 2 ⊕ · · · ⊕ v N ( ⊕ associative, non commutative) ◮ partial result of the Reduce operation: v [ k,m ] = v k ⊕ v 2 ⊕ · · · ⊕ v m ◮ two partial results can be merged: v [ k,m ] = v [ k,l ] ⊕ v [ l +1 ,m ] (computational task T k,l,m ) Loris Marchal Steady state collective communications 17/ 27

  53. Reduce - Reduction trees ◮ Reduce: ◮ each processor P r i owns a value v i ◮ compute V = v 1 ⊕ v 2 ⊕ · · · ⊕ v N ( ⊕ associative, non commutative) ◮ partial result of the Reduce operation: v [ k,m ] = v k ⊕ v 2 ⊕ · · · ⊕ v m ◮ two partial results can be merged: v [ k,m ] = v [ k,l ] ⊕ v [ l +1 ,m ] (computational task T k,l,m ) Loris Marchal Steady state collective communications 17/ 27

  54. Reduce - Reduction trees ◮ Reduce: ◮ each processor P r i owns a value v i ◮ compute V = v 1 ⊕ v 2 ⊕ · · · ⊕ v N ( ⊕ associative, non commutative) ◮ partial result of the Reduce operation: v [ k,m ] = v k ⊕ v 2 ⊕ · · · ⊕ v m ◮ two partial results can be merged: v [ k,m ] = v [ k,l ] ⊕ v [ l +1 ,m ] (computational task T k,l,m ) Loris Marchal Steady state collective communications 17/ 27

  55. Reduce - Reduction trees ◮ Reduce: ◮ each processor P r i owns a value v i ◮ compute V = v 1 ⊕ v 2 ⊕ · · · ⊕ v N ( ⊕ associative, non commutative) ◮ partial result of the Reduce operation: v [ k,m ] = v k ⊕ v 2 ⊕ · · · ⊕ v m ◮ two partial results can be merged: v [ k,m ] = v [ k,l ] ⊕ v [ l +1 ,m ] (computational task T k,l,m ) Loris Marchal Steady state collective communications 17/ 27

  56. Reduce - Reduction trees ◮ Reduce: ◮ each processor P r i owns a value v i ◮ compute V = v 1 ⊕ v 2 ⊕ · · · ⊕ v N ( ⊕ associative, non commutative) ◮ partial result of the Reduce operation: v [ k,m ] = v k ⊕ v 2 ⊕ · · · ⊕ v m ◮ two partial results can be merged: v [ k,m ] = v [ k,l ] ⊕ v [ l +1 ,m ] (computational task T k,l,m ) Loris Marchal Steady state collective communications 17/ 27

  57. Reduce - Reduction trees v 0 v 1 v 2 ◮ Reduce: P 0 P 1 P 2 ◮ each processor P r i owns a value v i ◮ compute V = v 1 ⊕ v 2 ⊕ · · · ⊕ v N ( ⊕ associative, non commutative) ◮ partial result of the Reduce operation: v [ k,m ] = v k ⊕ v 2 ⊕ · · · ⊕ v m ◮ two partial results can be merged: v [ k,m ] = v [ k,l ] ⊕ v [ l +1 ,m ] (computational task T k,l,m ) Loris Marchal Steady state collective communications 17/ 27

  58. Reduce - Reduction trees v 0 v 1 v 2 ◮ Reduce: P 0 P 1 P 2 ◮ each processor P r i owns a value v i v 2 ◮ compute V = v 1 ⊕ v 2 ⊕ · · · ⊕ v N P 2 → P 1 ( ⊕ associative, non commutative) ◮ partial result of the Reduce operation: v [ k,m ] = v k ⊕ v 2 ⊕ · · · ⊕ v m ◮ two partial results can be merged: v [ k,m ] = v [ k,l ] ⊕ v [ l +1 ,m ] (computational task T k,l,m ) Loris Marchal Steady state collective communications 17/ 27

  59. Reduce - Reduction trees v 0 v 1 v 2 ◮ Reduce: P 0 P 1 P 2 ◮ each processor P r i owns a value v i v 2 ◮ compute V = v 1 ⊕ v 2 ⊕ · · · ⊕ v N P 2 → P 1 ( ⊕ associative, non commutative) T 1 , 1 , 2 ◮ partial result of the Reduce P 1 operation: v [ k,m ] = v k ⊕ v 2 ⊕ · · · ⊕ v m ◮ two partial results can be merged: v [ k,m ] = v [ k,l ] ⊕ v [ l +1 ,m ] (computational task T k,l,m ) Loris Marchal Steady state collective communications 17/ 27

  60. Reduce - Reduction trees v 0 v 1 v 2 ◮ Reduce: P 0 P 1 P 2 ◮ each processor P r i owns a value v i v 2 ◮ compute V = v 1 ⊕ v 2 ⊕ · · · ⊕ v N P 2 → P 1 ( ⊕ associative, non commutative) v 0 T 1 , 1 , 2 ◮ partial result of the Reduce P 0 → P 1 P 1 operation: v [ k,m ] = v k ⊕ v 2 ⊕ · · · ⊕ v m ◮ two partial results can be merged: v [ k,m ] = v [ k,l ] ⊕ v [ l +1 ,m ] (computational task T k,l,m ) Loris Marchal Steady state collective communications 17/ 27

  61. Reduce - Reduction trees v 0 v 1 v 2 ◮ Reduce: P 0 P 1 P 2 ◮ each processor P r i owns a value v i v 2 ◮ compute V = v 1 ⊕ v 2 ⊕ · · · ⊕ v N P 2 → P 1 ( ⊕ associative, non commutative) v 0 T 1 , 1 , 2 ◮ partial result of the Reduce P 0 → P 1 P 1 operation: v [ k,m ] = v k ⊕ v 2 ⊕ · · · ⊕ v m T 0 , 0 , 2 P 1 ◮ two partial results can be merged: v [ k,m ] = v [ k,l ] ⊕ v [ l +1 ,m ] (computational task T k,l,m ) Loris Marchal Steady state collective communications 17/ 27

  62. Reduce - Reduction trees v 0 v 1 v 2 ◮ Reduce: P 0 P 1 P 2 ◮ each processor P r i owns a value v i v 2 ◮ compute V = v 1 ⊕ v 2 ⊕ · · · ⊕ v N P 2 → P 1 ( ⊕ associative, non commutative) v 0 T 1 , 1 , 2 ◮ partial result of the Reduce P 0 → P 1 P 1 operation: v [ k,m ] = v k ⊕ v 2 ⊕ · · · ⊕ v m T 0 , 0 , 2 P 1 ◮ two partial results can be merged: v [ k,m ] = v [ k,l ] ⊕ v [ l +1 ,m ] v [0 , 2] (computational task T k,l,m ) P 1 → P 0 Loris Marchal Steady state collective communications 17/ 27

  63. Series of Reduce ◮ each processor P r i owns a set of values v t i (e.g. produced at different time-steps t ) ◮ perform a Reduce operation on each set { v t 1 , . . . , v t N } to compute V t ◮ each reduction uses a reduction tree ◮ two reductions ( t 1 and t 2 ) may use different trees Loris Marchal Steady state collective communications 18/ 27

  64. Linear Program - Notations ◮ s ( P i → P j , v [ k,l ] ) : fractional number of values v [ k,l ] sent on link P i → P j within one time-unit ◮ t ( P i → P j ) fractional occupation time of link P i → P j within one time-unit: 0 � t ( P i → P j ) � 1 ◮ cons ( P i , T k,l,m ) : fractional number of tasks T k,l,m computed on processor P i within one time-unit ◮ α ( P i ) time spent by processor P i computing tasks within one time-unit: 0 � α ( P i ) � 1 ◮ size ( v [ k,m ] ) size of a message containing a value v t [ k,m ] ◮ w ( P i , T k,l,m ) time needed by processor P i to compute one task T k,l,m Loris Marchal Steady state collective communications 19/ 27

  65. Linear Program - Notations ◮ s ( P i → P j , v [ k,l ] ) : fractional number of values v [ k,l ] sent on link P i → P j within one time-unit ◮ t ( P i → P j ) fractional occupation time of link P i → P j within one time-unit: 0 � t ( P i → P j ) � 1 ◮ cons ( P i , T k,l,m ) : fractional number of tasks T k,l,m computed on processor P i within one time-unit ◮ α ( P i ) time spent by processor P i computing tasks within one time-unit: 0 � α ( P i ) � 1 ◮ size ( v [ k,m ] ) size of a message containing a value v t [ k,m ] ◮ w ( P i , T k,l,m ) time needed by processor P i to compute one task T k,l,m Loris Marchal Steady state collective communications 19/ 27

  66. Linear Program - Notations ◮ s ( P i → P j , v [ k,l ] ) : fractional number of values v [ k,l ] sent on link P i → P j within one time-unit ◮ t ( P i → P j ) fractional occupation time of link P i → P j within one time-unit: 0 � t ( P i → P j ) � 1 ◮ cons ( P i , T k,l,m ) : fractional number of tasks T k,l,m computed on processor P i within one time-unit ◮ α ( P i ) time spent by processor P i computing tasks within one time-unit: 0 � α ( P i ) � 1 ◮ size ( v [ k,m ] ) size of a message containing a value v t [ k,m ] ◮ w ( P i , T k,l,m ) time needed by processor P i to compute one task T k,l,m Loris Marchal Steady state collective communications 19/ 27

  67. Linear Program - Notations ◮ s ( P i → P j , v [ k,l ] ) : fractional number of values v [ k,l ] sent on link P i → P j within one time-unit ◮ t ( P i → P j ) fractional occupation time of link P i → P j within one time-unit: 0 � t ( P i → P j ) � 1 ◮ cons ( P i , T k,l,m ) : fractional number of tasks T k,l,m computed on processor P i within one time-unit ◮ α ( P i ) time spent by processor P i computing tasks within one time-unit: 0 � α ( P i ) � 1 ◮ size ( v [ k,m ] ) size of a message containing a value v t [ k,m ] ◮ w ( P i , T k,l,m ) time needed by processor P i to compute one task T k,l,m Loris Marchal Steady state collective communications 19/ 27

  68. Linear Program - Notations ◮ s ( P i → P j , v [ k,l ] ) : fractional number of values v [ k,l ] sent on link P i → P j within one time-unit ◮ t ( P i → P j ) fractional occupation time of link P i → P j within one time-unit: 0 � t ( P i → P j ) � 1 ◮ cons ( P i , T k,l,m ) : fractional number of tasks T k,l,m computed on processor P i within one time-unit ◮ α ( P i ) time spent by processor P i computing tasks within one time-unit: 0 � α ( P i ) � 1 ◮ size ( v [ k,m ] ) size of a message containing a value v t [ k,m ] ◮ w ( P i , T k,l,m ) time needed by processor P i to compute one task T k,l,m Loris Marchal Steady state collective communications 19/ 27

  69. Linear Program - Notations ◮ s ( P i → P j , v [ k,l ] ) : fractional number of values v [ k,l ] sent on link P i → P j within one time-unit ◮ t ( P i → P j ) fractional occupation time of link P i → P j within one time-unit: 0 � t ( P i → P j ) � 1 ◮ cons ( P i , T k,l,m ) : fractional number of tasks T k,l,m computed on processor P i within one time-unit ◮ α ( P i ) time spent by processor P i computing tasks within one time-unit: 0 � α ( P i ) � 1 ◮ size ( v [ k,m ] ) size of a message containing a value v t [ k,m ] ◮ w ( P i , T k,l,m ) time needed by processor P i to compute one task T k,l,m Loris Marchal Steady state collective communications 19/ 27

  70. Linear Program - Constraints ◮ occupation of a link P i → P j : � t ( P i → P j ) = s ( P i → P j , v [ k,l ] ) × size ( v [ k,l ] ) × c ( i, j ) v [ k,l ] ◮ occupation time of a processor P i : � α ( P i ) = cons ( P i , T k,l,m ) × w ( P i , T k,l,m ) T k,l,m ◮ “conservation law” for packets of type v [ k,m ] : � � s ( P j → P i , v [ k,m ] ) + cons ( P i , T k,l,m ) P j → P i k � l<m � � � = s ( P i → P j , v [ k,m ] ) + cons ( P i , T k,m,n ) + cons ( P i , T n,k − 1 ,m ) n>m P i → P j n<k Loris Marchal Steady state collective communications 20/ 27

  71. Linear Program - Constraints ◮ occupation of a link P i → P j : � t ( P i → P j ) = s ( P i → P j , v [ k,l ] ) × size ( v [ k,l ] ) × c ( i, j ) v [ k,l ] ◮ occupation time of a processor P i : � α ( P i ) = cons ( P i , T k,l,m ) × w ( P i , T k,l,m ) T k,l,m ◮ “conservation law” for packets of type v [ k,m ] : � � s ( P j → P i , v [ k,m ] ) + cons ( P i , T k,l,m ) P j → P i k � l<m � � � = s ( P i → P j , v [ k,m ] ) + cons ( P i , T k,m,n ) + cons ( P i , T n,k − 1 ,m ) n>m P i → P j n<k Loris Marchal Steady state collective communications 20/ 27

  72. Linear Program - Constraints ◮ occupation of a link P i → P j : � t ( P i → P j ) = s ( P i → P j , v [ k,l ] ) × size ( v [ k,l ] ) × c ( i, j ) v [ k,l ] ◮ occupation time of a processor P i : � α ( P i ) = cons ( P i , T k,l,m ) × w ( P i , T k,l,m ) T k,l,m ◮ “conservation law” for packets of type v [ k,m ] : � � s ( P j → P i , v [ k,m ] ) + cons ( P i , T k,l,m ) P j → P i k � l<m � � � = s ( P i → P j , v [ k,m ] ) + cons ( P i , T k,m,n ) + cons ( P i , T n,k − 1 ,m ) n>m P i → P j n<k Loris Marchal Steady state collective communications 20/ 27

  73. Linear Program - Constraints ◮ definition of the throughput: � � TP = s ( P j → P target , v [0 ,m ] )+ cons ( P target , T 0 ,l,N ) P j → P target 0 � l<N − 1 ◮ solve the following linear program over the rational numbers: Steady-State Reduce Problem on a Graph SSRP(G) Maximize TP , subject to all previous constraints Loris Marchal Steady state collective communications 21/ 27

  74. Linear Program - Constraints ◮ definition of the throughput: � � TP = s ( P j → P target , v [0 ,m ] )+ cons ( P target , T 0 ,l,N ) P j → P target 0 � l<N − 1 ◮ solve the following linear program over the rational numbers: Steady-State Reduce Problem on a Graph SSRP(G) Maximize TP , subject to all previous constraints Loris Marchal Steady state collective communications 21/ 27

  75. Building a schedule ◮ consider the reduction tree T t associated with the computation of the t th value ( V t ): ◮ a given tree may be used by many time-stamps t ◮ there exists an algorithm which extracts from the solution a set of weighted trees such that ◮ this description is polynomial and ◮ the sum of the weighted trees is equal to the original solution ◮ same use of a weighted edge-coloring algorithm on a bipartite graph to orchestrate the communication Loris Marchal Steady state collective communications 22/ 27

  76. Building a schedule ◮ consider the reduction tree T t associated with the computation of the t th value ( V t ): ◮ a given tree may be used by many time-stamps t ◮ there exists an algorithm which extracts from the solution a set of weighted trees such that ◮ this description is polynomial and ◮ the sum of the weighted trees is equal to the original solution ◮ same use of a weighted edge-coloring algorithm on a bipartite graph to orchestrate the communication Loris Marchal Steady state collective communications 22/ 27

  77. Building a schedule ◮ consider the reduction tree T t associated with the computation of the t th value ( V t ): ◮ a given tree may be used by many time-stamps t ◮ there exists an algorithm which extracts from the solution a set of weighted trees such that ◮ this description is polynomial and ◮ the sum of the weighted trees is equal to the original solution ◮ same use of a weighted edge-coloring algorithm on a bipartite graph to orchestrate the communication Loris Marchal Steady state collective communications 22/ 27

  78. Building a schedule ◮ consider the reduction tree T t associated with the computation of the t th value ( V t ): ◮ a given tree may be used by many time-stamps t ◮ there exists an algorithm which extracts from the solution a set of weighted trees such that ◮ this description is polynomial and ◮ the sum of the weighted trees is equal to the original solution ◮ same use of a weighted edge-coloring algorithm on a bipartite graph to orchestrate the communication Loris Marchal Steady state collective communications 22/ 27

  79. Building a schedule ◮ consider the reduction tree T t associated with the computation of the t th value ( V t ): ◮ a given tree may be used by many time-stamps t ◮ there exists an algorithm which extracts from the solution a set of weighted trees such that ◮ this description is polynomial and ◮ the sum of the weighted trees is equal to the original solution ◮ same use of a weighted edge-coloring algorithm on a bipartite graph to orchestrate the communication Loris Marchal Steady state collective communications 22/ 27

  80. Building a schedule ◮ consider the reduction tree T t associated with the computation of the t th value ( V t ): ◮ a given tree may be used by many time-stamps t ◮ there exists an algorithm which extracts from the solution a set of weighted trees such that ◮ this description is polynomial and ◮ the sum of the weighted trees is equal to the original solution ◮ same use of a weighted edge-coloring algorithm on a bipartite graph to orchestrate the communication Loris Marchal Steady state collective communications 22/ 27

  81. Toy Example for Series of Reduce 0 1 1 1 1 2 1 topology Loris Marchal Steady state collective communications 23/ 27

  82. Toy Example for Series of Reduce 1 T 0 , 0 , 2 0 2 1 3 v [1 , 2] 3 v [1 , 2] 2 3 v [1 , 1] 1 2 1 3 T 1 , 1 , 2 2 3 T 1 , 1 , 2 1 3 v [2 , 2] results of the linear program Loris Marchal Steady state collective communications 23/ 27

Recommend

Pandemic Telehealth vs Steady-State Telehealth In steady-state telepractice, several

Pandemic Telehealth vs Steady-State Telehealth In steady-state telepractice, several

Pandemic Telehealth vs Steady-State Telehealth In steady-state telepractice, several considerations are taken prior to starting therapy to determine appropriateness of the model. Equipment Client preference for therapy modes (in-person

459 views • 7 slides
Steady State Equation for Wells After a well has been pumped for an extended period, steady

Steady State Equation for Wells After a well has been pumped for an extended period, steady

Steady State Equation for Wells After a well has been pumped for an extended period, steady state conditions are approached In the simplest possible case, use Darcys equation for an isotropic, homogeneous aquifer a) Confined

459 views • 12 slides
Steady-State Simulation Steady-State Simulation Overview Reading: Ch. 9 in Law &amp; Ch. 15 in

Steady-State Simulation Steady-State Simulation Overview Reading: Ch. 9 in Law &amp; Ch. 15 in

Steady-State Simulation Steady-State Simulation Overview Reading: Ch. 9 in Law &amp; Ch. 15 in Handbook of Simulation The Regenerative Method Regenerative processes Regenerative Simulation Peter J. Haas Delays The Batch Means Method

743 views • 9 slides
Optimizing Phenotypic HBV Drug Resistance Testing for High Throughput Ke Zhang, Rolf Kaiser,

Optimizing Phenotypic HBV Drug Resistance Testing for High Throughput Ke Zhang, Rolf Kaiser,

Optimizing Phenotypic HBV Drug Resistance Testing for High Throughput Ke Zhang, Rolf Kaiser, Ulrike Protzer Institute of Virology Treatment Options for Chronic Hepatitis B Approved for HBV Unlabeled Investigational Interferon alfa-2b

296 views • 15 slides
1 Outline Part I: Theory steady state enzyme kinetics: a new approach Part II: Experiment

1 Outline Part I: Theory steady state enzyme kinetics: a new approach Part II: Experiment

A New 'Microscopic' Look at Steady-state Enzyme Kinetics Petr Kuzmi BioKin Ltd. http://www.biokin.com SEMINAR: University of Massachusetts Medical School Worcester, MA April 6, 2015 Steady-State Enzyme Kinetics 1 1 Outline Part I:

648 views • 36 slides
Optimizing Throughput with Network Coding Zongpeng Li, Baochun Li Department of Electrical and

Optimizing Throughput with Network Coding Zongpeng Li, Baochun Li Department of Electrical and

Optimizing Throughput with Network Coding Zongpeng Li, Baochun Li Department of Electrical and Computer Engineering University of T oronto DIMACS W orking Group on Network Coding Rutgers University January 28, 2005 URL: google Baochun

722 views • 55 slides
Outline Server Utilization System Performance Steady-State Behavior of

Outline Server Utilization System Performance Steady-State Behavior of

Chapter 6 Queueing Models (2) Banks, Carson, Nelson &amp; Nicol Discrete-Event System Simulation Outline Server Utilization System Performance Steady-State Behavior of Infinite-Population Models Steady-State Behavior of

286 views • 26 slides
Steady-State Simulation Reading: Ch. 9 in Law &amp; Ch. 15 in Handbook of Simulation Peter J.

Steady-State Simulation Reading: Ch. 9 in Law &amp; Ch. 15 in Handbook of Simulation Peter J.

Steady-State Simulation Reading: Ch. 9 in Law &amp; Ch. 15 in Handbook of Simulation Peter J. Haas CS 590M: Simulation Spring Semester 2020 1 / 34 Steady-State Simulation Overview The Regenerative Method Regenerative processes Regenerative

345 views • 34 slides
MPI types, Scatter and Scatterv MPI types, Scatter and Scatterv 0 1 2 3 4 5 Logical and

MPI types, Scatter and Scatterv MPI types, Scatter and Scatterv 0 1 2 3 4 5 Logical and

MPI types, Scatter and Scatterv MPI types, Scatter and Scatterv 0 1 2 3 4 5 Logical and physical 6 7 8 9 10 11 layout of a C/C++ array in 12 13 14 15 16 17 memory. 18 19 20 21 22 23 A = malloc(6*6*sizeof(int)); 24 25 26 27 28 29

635 views • 42 slides
PHI: ARCHITECTURAL SUPPORT FOR SYNCHRONIZATION- AND BANDWIDTH-EFFICIENT COMMUTATIVE SCATTER

PHI: ARCHITECTURAL SUPPORT FOR SYNCHRONIZATION- AND BANDWIDTH-EFFICIENT COMMUTATIVE SCATTER

PHI: ARCHITECTURAL SUPPORT FOR SYNCHRONIZATION- AND BANDWIDTH-EFFICIENT COMMUTATIVE SCATTER UPDATES Anurag Mukkara , Nathan Beckmann, Daniel Sanchez MICRO 2019 Scatter updates are common but inefficient 2 Scatter updates are common but

1.55k views • 152 slides
Steady State Temperature Steady State Temperature Profiles in Rods Profiles in Rods Amy Chan

Steady State Temperature Steady State Temperature Profiles in Rods Profiles in Rods Amy Chan

Steady State Temperature Steady State Temperature Profiles in Rods Profiles in Rods Amy Chan Amy Chan Anders Berliner, Catherine Chau Chau, , Erise Erise Hosoya Hosoya Anders Berliner, Catherine Case Western Reserve University Case

361 views • 16 slides
Steady State Property Verification: a Comparison Study Diana EL RABIH ( 1 ) , Gael Gorgo ( 2 ) ,

Steady State Property Verification: a Comparison Study Diana EL RABIH ( 1 ) , Gael Gorgo ( 2 ) ,

Steady State Property Verification: a Comparison Study Steady State Property Verification: a Comparison Study Diana EL RABIH ( 1 ) , Gael Gorgo ( 2 ) , Nihal PEKERGIN ( 1 ) , Jean-Marc Vincent ( 2 ) ( 1 ) LACL, University of Paris Est (Paris 12) (

992 views • 68 slides
1 Enzyme kinetic modeling and its importance WHAT CAN ENZYME KINETICS DO FOR US? macroscopic

1 Enzyme kinetic modeling and its importance WHAT CAN ENZYME KINETICS DO FOR US? macroscopic

A New 'Microscopic' Look at Steady-state Enzyme Kinetics Petr Kuzmi BioKin Ltd. http://www.biokin.com SEMINAR: University of Massachusetts Medical School Worcester, MA April 6, 2015 Steady-State Enzyme Kinetics 1 Outline Part I: Theory

402 views • 18 slides
Plan of the Lecture Review: transient and steady-state response; DC gain and the FVT

Plan of the Lecture Review: transient and steady-state response; DC gain and the FVT

Plan of the Lecture Review: transient and steady-state response; DC gain and the FVT Todays topic: system-modeling diagrams; prototype 2nd-order system Plan of the Lecture Review: transient and steady-state response; DC gain and

1.18k views • 106 slides
high N , steady-state scenarios 5 W.W. (Bill) Heidbrink 1 Amplitude with J. Ferron, 2 C.

high N , steady-state scenarios 5 W.W. (Bill) Heidbrink 1 Amplitude with J. Ferron, 2 C.

Alfvn eigenmodes (AE) degrade fast-ion confinement in high N , steady-state scenarios 5 W.W. (Bill) Heidbrink 1 Amplitude with J. Ferron, 2 C. Holcomb, 3 M. Van Zeeland 2 , E. Bass 4 , X. Chen 2 , C. Collins 1 , A. Garofalo 2 , X. Gong 5 ,

540 views • 34 slides
Comparison of Measurement and Analysis of High Temperature Steady-state Creep Experiment for

Comparison of Measurement and Analysis of High Temperature Steady-state Creep Experiment for

Transactions of the Korean Nuclear Society Virtual Spring Meeting July 9-10, 2020 Comparison of Measurement and Analysis of High Temperature Steady-state Creep Experiment for Equipment Verification Gyeong-Ha Choi a,b , Dong-Hyun Kim a , JaeYong

306 views • 3 slides
Network Inference Using Steady- State Data and Goldbeter- Koshland Kinetics C.J.Oates 1 , 2 ,

Network Inference Using Steady- State Data and Goldbeter- Koshland Kinetics C.J.Oates 1 , 2 ,

Network Inference Using Steady- State Data and Goldbeter- Koshland Kinetics C.J.Oates 1 , 2 , B.T.Hennessy 3 , Y.Lu 4 , G.B.Mills 4 and S.Mukherjee 2 , 1 1 Dept. Biochemistry, Nederlands Kanker Instituut 2 Depts. Statistics &amp; Complexity, Univ.

805 views • 65 slides
Use of parallel matrix algorithms for Laplace partial differential equations A steady-state

Use of parallel matrix algorithms for Laplace partial differential equations A steady-state

CS140 V-1 Use of parallel matrix algorithms for Laplace partial differential equations A steady-state heat-flow problem on a rectangular 10 cm 20 cm metal sheet. One edge maintains temperature of 100 degree, other three edges

401 views • 18 slides
virtual memory 5 / devices 1 last time page replacement metrics optimizing hit rate really

virtual memory 5 / devices 1 last time page replacement metrics optimizing hit rate really

virtual memory 5 / devices 1 last time page replacement metrics optimizing hit rate really care about throughput other possibilities (like processor scheduling) Beladys MIN: ideal hit rate policy replace what is accessed furthest in

942 views • 79 slides
Dynamic fmuid connectjvity during steady- state multjphase fmow in a sandstone Catriona Reynolds,

Dynamic fmuid connectjvity during steady- state multjphase fmow in a sandstone Catriona Reynolds,

Dynamic fmuid connectjvity during steady- state multjphase fmow in a sandstone Catriona Reynolds, Hannah Menke, Matuhew Andrew, Martjn Blunt &amp; Sam Krevor Proceedings of the Natjonal Academy of Science (2017) 114:31, 8187-8192 Department of

672 views • 39 slides
Steady-State Scheduling on Heterogeneous Platforms Matthieu Gallet Advisors: Yves Robert and

Steady-State Scheduling on Heterogeneous Platforms Matthieu Gallet Advisors: Yves Robert and

tel-00637362, version 1 - 1 Nov 2011 Steady-State Scheduling on Heterogeneous Platforms Matthieu Gallet Advisors: Yves Robert and Fr ed eric Vivien Ecole Normale Sup erieure de Lyon GRAAL team, Laboratoire de lInformatique du

1.54k views • 142 slides
Solving the steady state diffusion equation with uncertainty Mid-year presentation Virginia

Solving the steady state diffusion equation with uncertainty Mid-year presentation Virginia

Background Methods Results Looking ahead References Solving the steady state diffusion equation with uncertainty Mid-year presentation Virginia Forstall vhfors@gmail.com Advisor: Howard Elman elman@cs.umd.edu Department of Computer

801 views • 52 slides
Dynamic and Steady-State Process Investigations Using Functional Data Analysis P. James McLellan

Dynamic and Steady-State Process Investigations Using Functional Data Analysis P. James McLellan

Dynamic and Steady-State Process Investigations Using Functional Data Analysis P. James McLellan Department of Chemical Engineering Queens University Kingston, ON Canada mclellnj@chee.queensu.ca 1 Outline Brief Exercise

801 views • 60 slides
Unsteady Vehicle Simulation Why run unsteady? Steady State Advantages Quick turn-around

Unsteady Vehicle Simulation Why run unsteady? Steady State Advantages Quick turn-around

Unsteady Vehicle Simulation Why run unsteady? Steady State Advantages Quick turn-around time Proven accuracy in absolute values, and trends Excellent tool for optimization studies to reduce drag. Proven to provide as accurate

406 views • 28 slides

More recommend