A PCIe Congestion-Aware Performance Model for Densely Populated Accelerator Servers

Transcription

1 A PCIe Congestion-Aware Performance Model for Densely Populated Accelerator Servers Maxime Martinasso, Grzegorz Kwasniewski, Sadaf R. Alam, Thomas C. Schulthess, Torsten Hoefler Swiss National Supercomputing Centre, ETH Zurich, 6900 Lugano, Switzerland Department of Computer Science, ETH Zurich, Universitätstr. 6, 8092 Zurich, Switzerland Institute for Theoretical Physics, ETH Zurich, 8093 Zurich, Switzerland Computer Science and Mathematics Division, Oak Ridge National Laboratory, USA

2 Why more densely populated accelerator servers? accelerators are faster and more energy-efficient than CPU densely populated accelerator servers are high performance nodes reduce space occupancy of the data center HPC Advisory Council A PCIe performance model 2

3 Cray CS Storm new MeteoSwiss supercomputer 2 cabinets 12 hybrid computing nodes per cabinet 2 Intel Haswell 12-core CPUs per node 8 NVIDIA Tesla K80 GPU accelerators per node 2 GPU processors per accelerator 192 GPU processors in total 360 GPU teraflops in total Production system GPUs connected by PCI-Express HPC Advisory Council A PCIe performance model 3

4 generation 3, 16 GB/s using x16 wide lane dual simplex (a pair of unidirectional links) exchange buffer availability between pair of ports of a link tree-based topology Building a densely populated accelerator servers with PCIe: CPU CPU Root Complex Legend: PCIe Root Complex PCIe Switch GPU Port x16 Link HPC Advisory Council A PCIe performance model 4

5 generation 3, 16 GB/s using x16 wide lane dual simplex (a pair of unidirectional links) exchange buffer availability between pair of ports of a link tree-based topology Building a densely populated accelerator servers with PCIe: CPU CPU Root Complex GPU K80 Legend: PCIe Root Complex PCIe Switch GPU Port x16 Link HPC Advisory Council A PCIe performance model 4

6 generation 3, 16 GB/s using x16 wide lane dual simplex (a pair of unidirectional links) exchange buffer availability between pair of ports of a link tree-based topology Building a densely populated accelerator servers with PCIe: CPU CPU Root Complex GPU K80 GPU K80 Legend: PCIe Root Complex PCIe Switch GPU Port x16 Link HPC Advisory Council A PCIe performance model 4

7 generation 3, 16 GB/s using x16 wide lane dual simplex (a pair of unidirectional links) exchange buffer availability between pair of ports of a link tree-based topology Building a densely populated accelerator servers with PCIe: CPU CPU Root Complex GPU K80 Legend: PCIe Root Complex PCIe Switch GPU Port x16 Link GPU K80 GPU K80 HPC Advisory Council A PCIe performance model 4

8 generation 3, 16 GB/s using x16 wide lane dual simplex (a pair of unidirectional links) exchange buffer availability between pair of ports of a link tree-based topology Building a densely populated accelerator servers with PCIe: CPU CPU Root Complex Legend: PCIe Root Complex PCIe Switch GPU Port x16 Link GPU K80 GPU K80 GPU K80 GPU K80 HPC Advisory Council A PCIe performance model 4

9 generation 3, 16 GB/s using x16 wide lane dual simplex (a pair of unidirectional links) exchange buffer availability between pair of ports of a link tree-based topology Building a densely populated accelerator servers with PCIe: CPU CPU Root Complex Legend: PCIe Root Complex PCIe Switch GPU Port x16 Link K80 HPC Advisory Council A PCIe performance model 4

10 Communication conflicts 0 7, 1 4, HPC Advisory Council A PCIe performance model 5

11 Communication conflicts 0 7, 1 4, Upstream port conflict HPC Advisory Council A PCIe performance model 5

12 Communication conflicts 0 7, 1 4, Upstream port conflict Downstream port conflict HPC Advisory Council A PCIe performance model 5

13 Communication conflicts 0 7, 1 4, Crossing the Root Complex conflict 19 Upstream port conflict Downstream port conflict HPC Advisory Council A PCIe performance model 5

14 Communication conflicts 0 7, 1 4, Crossing the Root Complex conflict 19 Head-of-Line blocking Upstream port conflict Downstream port conflict HPC Advisory Council A PCIe performance model 5

15 18 Crossing the Root Complex conflict 19 Head-of-Line blocking Upstream port conflict Downstream port conflict Conflicts identify by: Carefully reading the PCIe and vendor documentation Self experiments with micro-benchmarks (looking for HoL Blocking) HPC Advisory Council A PCIe performance model 5

16 Motivation COSMO halo exchange GPU4 GPU5 GPU6 GPU7 GPU6 GPU7 GPU4 GPU5 GPU0 GPU1 GPU2 GPU3 2D domain decomposition GPU2 GPU3 GPU0 GPU1 3D domain decomposition Which order of communications is the fastest? HPC Advisory Council A PCIe performance model 6

17 Motivation COSMO halo exchange GPU4 GPU5 GPU6 GPU7 GPU6 GPU7 GPU4 GPU5 GPU0 GPU1 GPU2 GPU3 2D domain decomposition GPU2 GPU3 GPU0 GPU1 3D domain decomposition Which order of communications is the fastest? D domain decomposition example: 20, 376 possibilities 3D domain decomposition has more than 1.6 Million possibilities HPC Advisory Council A PCIe performance model 6

18 PCIe performance model We want to identify the congestion factors ρ [0, 1] which limit the available bandwidth per communication at each communication phase. model: compute Step (A) Step (B) Step (C) Step (D) Communication graph Source Arbitration Upstream port conflicts Downstream port conflicts + Head-of-Line Blocking Update messages Advance to next time step Update communication graph Yes Remaining messages? No End HPC Advisory Council A PCIe performance model 7

19 Communication phase update messages Elapsed time L c, message size M c, set of communication phases S c : Time L C0 = t 0 = 1 B m0,0 ρ 0,0 and M C0 = m 0,0 HPC Advisory Council A PCIe performance model 8

20 Communication phase update messages Elapsed time L c, message size M c, set of communication phases S c : Time L C0 = t 0 + t 1 = 1 B ( m0,0 ρ 0,0 + m0,1 ρ 0,1 ) and M C0 = m 0,0 + m 0,1 HPC Advisory Council A PCIe performance model 8

21 Communication phase update messages Elapsed time L c, message size M c, set of communication phases S c : Time L c = i S c t i = 1 B i S c m c,i ρ c,i and M c = i S c m c,i (1) start times tstart i and M c are known (2) ρ c,i are given by the model with (1) and (2) L c are computable HPC Advisory Council A PCIe performance model 8

22 Model conflicts on switch We want to identify the congestion factors ρ [0, 1] which limit the available bandwidth per communication at each communication phase. Each communication enters a switch with a congestion factor ρ and leaves with a congestion factor ρ If i ρ i > 1 then an arbitration policy is required, ρ = ρ otherwise HPC Advisory Council A PCIe performance model 9

23 Upstream port conflict Proportional sharing of available bandwidth HPC Advisory Council A PCIe performance model 10

24 Downstream port conflict Round-robin policy Performance reduction for crossing the root complex C R set of communications crossing the root complex; n number of grouped communication sets; R congestion factor of a grouped communication set; τ congestion factor for crossing the root complex; if C R = then R = 1 n if C R then R = { min(max( 1 τ, 0), R) n min( 1 + τ, R) n otherwise if R contains comm. CR HPC Advisory Council A PCIe performance model 11

25 Complete example Step (A) Source Arbitration Step (B) Upstream port conflicts 18 Step (C) Downstream port conflicts + 19 Step (D) Head-of-Line Blocking comm. Step (A) Step (B) Step (C) Step (D) Update (a) 0 2 messages 1 1/2 1/2 3/10 3/10 (b) /2 3/10 3/10 3/10 (c) /2 1/2 7/10 (d) /10 7/10 7/ Congestion graph step comm. cong. factor data remaining elapsed time (a) 0 2 3/ MB 36 ms (b) 1 4 3/ MB 36 ms τ = 0.2 Message size: 300MB Bandwidth: 11.6 GB/s (c) 3 2 7/10 0 MB 36 ms (d) /10 0 MB 36 ms Congestion graph step 2 comm. cong. factor data remaining elapsed time (a) 0 2 1/2 0 MB 65 ms (b) 1 4 1/2 0 MB 65 ms HPC Advisory Council A PCIe performance model 12

26 Model Validation Architecture parameters: B = 11.6GB/s τ = ,259 graphs: non-isomorphic cudamemcpyasync Communication pattern: scatter, gather, all-to-all Entire set of graphs for subsets of GPUs Randomly generated 100K communications Message size: 300 MB Time no contention: T ref = 25.3ms 95% of communication are in range +/- 15% HPC Advisory Council A PCIe performance model 13

27 Back to the motivation COSMO halo exchange Time [s] Upper limit on time to solution, throughput approach Running mode one instance per socket (8 GPUs) Large domain size 256x256x80 per GPU One step triggers 312 halo exchanges Message size: 40 KB to 254 KB Uses MPI (3!) 4 (2!) 4 = 20, 736 communication graphs for 2D domain GPU0 current implemented graph in COSMO GPU4 GPU1 GPU5 GPU2 GPU6 GPU3 2D domain decomposition Congestion HPC graphs Advisory Council sorted A PCIe performance by elapsed model 14 time GPU7 GPU4 GPU0 GPU6 GPU2 1.6x 1.9x GPU5 GPU1 GPU7 GPU3 3D domain decomposition

28 Fastest schedule for 2D decomposition HPC Advisory Council A PCIe performance model 15

29 Fastest schedule for 2D decomposition HPC Advisory Council A PCIe performance model 15

30 Fastest schedule for 2D decomposition HPC Advisory Council A PCIe performance model 15

31 Fastest schedule for 2D decomposition HPC Advisory Council A PCIe performance model 15

32 Fastest schedule for 3D decomposition HPC Advisory Council A PCIe performance model 16

33 Fastest schedule for 3D decomposition HPC Advisory Council A PCIe performance model 16

34 Fastest schedule for 3D decomposition HPC Advisory Council A PCIe performance model 16

35 Fastest schedule for 3D decomposition HPC Advisory Council A PCIe performance model 16

36 COSMO improvement fastest schedule COSMO gain: 5.6% per halo exchange step, gain is limited by MPI 2-sided overhead, initial tests with 1-sided seems to clearly improve the performance. HPC Advisory Council A PCIe performance model 17

37 Conclusion - Latency not modeled - MPI 2-sided overhead not modeled (use one-sided in progress) + Captures all PCIe features including congestion + Simple model only 2 parameters (B and τ ) + Precise for large messages + Design of topology-aware algorithms + COSMO halo exchange performance gain HPC Advisory Council A PCIe performance model 18

38 Thank you for your attention.

39 12 Measured bandwitdh [GB/s] x 1.21x 1.80x 1.88x slower for slower for slower for conflict 3 conflict 2 conflict or 0 3 alone 4 1 alone 0 3 in conflict in conflict in conflict K 8K 32K 128K 512K 2M 8M 32M Message size [Byte] τ = 1 1 /1.21 HPC Advisory Council A PCIe performance model 19

40 Legend: PCIe Root Complex PCIe Switch GPU Port x16 Link Socket Socket K80 K80 TOPOLOGY T1 TOPOLOGY T2 HPC Advisory Council A PCIe performance model 20