the power of remote GPU virtualization

Transcription

1 Creating new horizons with rcuda: the power of remote virtualization Federico Silla Universitat Politècnica de València Spain

2 s are great! HPC Advisory Council Perth Conference /55

3 Basics of computing (single node) Basic behavior of CUDA HPC Advisory Council Perth Conference /55

4 Basics of computing (single node) s can only be used within the node they are attached to Basic behavior of CUDA HPC Advisory Council Perth Conference /55

5 Using s across the cluster A -enabled cluster is a set of independent self-contained nodes that share nothing among them: MPI is required for aggregating resources within the cluster (included s) node 1 node 2 node 3 node n Interconnection HPC Advisory Council Perth Conference /55

6 s are great! How can we make an even better usage of s? HPC Advisory Council Perth Conference /55

7 s are great! How can we make an even better usage of s? Which characteristics do we miss from s? HPC Advisory Council Perth Conference /55

8 Characteristics missing in s How can we make an even better usage of s? Which characteristics do we miss from s? 1. Many s in a single box 2. Easily sharing a given (or s) HPC Advisory Council Perth Conference /55

9 Characteristics missing in s 1. Why many s in a single box Traditionally, in order to use many s, applications had to use MPI: s can only be used within the node they are attached to Nothing is directly shared among nodes (MPI required for aggregating computing resources within the cluster) node 1 node 2 node 3 node n A non-mpi application running in this node can only use the s in this node Interconnection HPC Advisory Council Perth Conference /55

10 Characteristics missing in s 1. Many s in a single box HPC Advisory Council Perth Conference /55

11 Characteristics missing in s 1. Many s in a single box The amount of s is limited by the physical space inside the node HPC Advisory Council Perth Conference /55

12 Characteristics missing in s 1. Having many s in a single box MonteCarlo multi- program running in 10 NVIDIA Tesla K40 s HPC Advisory Council Perth Conference /55

13 Characteristics missing in s 1. Having many s in a single box 64 s!! HPC Advisory Council Perth Conference /55

14 Characteristics missing in s 1. Many s in a single box How many s are many s in a single box? HPC Advisory Council Perth Conference /55

15 Characteristics missing in s 1. Many s in a single box How many s are many s in a single box? As many s as they can be installed in the cluster node 1 node 2 node 3 node n Interconnection HPC Advisory Council Perth Conference /55

16 Characteristics missing in s How can we make an even better usage of s? Which characteristics do we miss from s? 1. Many s in a single box 2. Easily sharing a given (or s) HPC Advisory Council Perth Conference /55

17 Characteristics missing in s 2. Easily sharing a given Why should we be interested on sharing s among applications? HPC Advisory Council Perth Conference /55

18 usage of -Blast assigned but not used assigned but not used NVIDIA Tesla K20 HPC Advisory Council Perth Conference /55

19 usage of CUDA-MEME utilization is far away from maximum NVIDIA Tesla K20 HPC Advisory Council Perth Conference /55

20 usage of LAMMPS assigned but not used NVIDIA Tesla K20 HPC Advisory Council Perth Conference /55

21 Sharing a among jobs: -Blast One instance required about 51 seconds Two concurrent instances of -Blast HPC Advisory Council Perth Conference /55

22 Sharing a among jobs: -Blast First instance Two concurrent instances of -Blast HPC Advisory Council Perth Conference /55

23 Sharing a among jobs: -Blast Second instance First instance Two concurrent instances of -Blast HPC Advisory Council Perth Conference /55

24 Sharing a among applications K20 (5GB memory) LAMMPS: 876 MB mcuda-meme: 151 MB BarraCUDA: 3319 MB MUMmer: 2104 MB -LIBSVM: 145 MB HPC Advisory Council Perth Conference /55

25 Sharing a among applications K20 (5GB memory) LAMMPS: 876 MB mcuda-meme: 151 MB BarraCUDA: 3319 MB MUMmer: 2104 MB -LIBSVM: 145 MB The main concern for sharing a is the memory limitation HPC Advisory Council Perth Conference /55

26 Characteristics missing in s How can we make an even better usage of s? Which characteristics do we miss from s? 1. Many s in a single box 2. Easily sharing a given (or s) HPC Advisory Council Perth Conference /55

27 Easily sharing a among VMs A is assigned to a VM by using PCI passthrough Assignment is done exclusively to a single virtual machine. Concurrent usage of the is not possible HPC Advisory Council Perth Conference /55

28 Easily sharing a among VMs High performance network available Low performance network available HPC Advisory Council Perth Conference /55

29 Characteristics missing in s Which characteristics do we miss from s? 1. Many s in a single box 2. Easily sharing a given (or s) HPC Advisory Council Perth Conference /55

30 Characteristics missing in s Which characteristics do we miss from s? 1. Many s in a single box 2. Easily sharing a given (or s) The remote virtualization technique can efficiently address these concerns HPC Advisory Council Perth Conference /55

31 Characteristics missing in s node 1 node 2 node 3 node n Interconnection The remote virtualization technique can efficiently address these concerns HPC Advisory Council Perth Conference /55

32 Characteristics missing in s Interconnection The remote virtualization technique can efficiently address these concerns HPC Advisory Council Perth Conference /55

33 Characteristics missing in s Interconnection The remote virtualization technique can efficiently address these concerns HPC Advisory Council Perth Conference /55

34 Characteristics missing in s Interconnection The remote virtualization technique can efficiently address these concerns HPC Advisory Council Perth Conference /55

35 Remote virtualization What is remote virtualization? HPC Advisory Council Perth Conference /55

36 Basics of computing Basic behavior of CUDA HPC Advisory Council Perth Conference /55

37 Basics of computing HPC Advisory Council Perth Conference /55

38 Remote virtualization No HPC Advisory Council Perth Conference /55

39 rcuda remote CUDA A software technology that enables a more flexible use of s in computing facilities No rcuda is a development by Technical University of Valencia HPC Advisory Council Perth Conference /55

40 Basics of rcuda rcuda is a development by Universitat Politècnica de València, Spain HPC Advisory Council Perth Conference /55

41 Basics of rcuda rcuda is a development by Universitat Politècnica de València, Spain HPC Advisory Council Perth Conference /55

42 Remote virtualization envision Remote virtualization allows a new vision of a deployment, moving from the usual cluster configuration: node 1 node 2 node 3 node n Physical configuration Interconnection to the following one: node 1 Logical connections node 2 node 3 node n Logical configuration Interconnection HPC Advisory Council Perth Conference /55

43 Performance of rcuda Guy Kawasaki, marketing specialist and Silicon Valley venture capitalist Ideas Are Easy, Implementation Is Hard HPC Advisory Council Perth Conference /55

44 Performance of rcuda to Higher is better to HPC Advisory Council Perth Conference /55

45 Performance of rcuda to Higher is better to HPC Advisory Council Perth Conference /55

46 Performance of rcuda to Higher is better to HPC Advisory Council Perth Conference /55

47 Performance of rcuda CUDA rcuda rcuda scenario 1 rcuda scenario 2 HPC Advisory Council Perth Conference /55

48 Performance of rcuda CUDA rcuda rcuda scenario 1 Ideas Are Easy, Implementation Is Hard rcuda scenario 2 Guy Kawasaki, marketing specialist and Silicon Valley venture capitalist Higher is better HPC Advisory Council Perth Conference /55

49 Performance of applications using rcuda K40 s and EDR InfiniBand Lower is better MonteCarlo multi- program running in 10 NVIDIA Tesla K40 s HPC Advisory Council Perth Conference /55

50 Performance of applications using rcuda 64 s!! HPC Advisory Council Perth Conference /55

51 Performance of applications using rcuda K20 and FDR InfiniBand K40 and EDR InfiniBand Lower is better HPC Advisory Council Perth Conference /55

52 Performance of applications using rcuda EDR InfiniBand and P100 Lower is better BarraCUDA CUDA-MEME Lower is better HPC Advisory Council Perth Conference /55

53 Get a free copy of rcuda at More than 850 requests world rcuda is a development by Technical University of Valencia HPC Advisory Council Perth Conference /55

54 Tony Díaz Pablo Higueras Javier Prades Carlos Reaño Jaime Sierra Federico Silla rcuda is a development by Technical University of Valencia HPC Advisory Council Perth Conference /55

55 Thanks! Questions? rcuda is a development by Technical University of Valencia HPC Advisory Council Perth Conference /55