The Next Frontier of Cloud Computing is in the Clouds, Literally

Transcription

1 The Next Frontier of Cloud Computing is in the Clouds, Literally Silviu Craciunas, Andreas Haas Christoph Kirsch, Hannes Payer Harald Röck, Andreas Rottmann Ana Sokolova, Rainer Trummer Joshua Love Raja Sengupta Universität Salzburg UC Berkeley Cornell University, AI-Sys-Robo Seminar, February 2011

2 The JAviator javiator.cs.uni-salzburg.at

3 Quad-Rotor Helicopter all carbon, titanium, aluminum design custom motors 1.3m diameter ~2.2kg weight +2kg payload [AIAA GNC 2008] ~40min (empty) ~10min (full)

4 Open Source Blueprints

5 Minimal # of Different Parts

6 Minimal # of Different Parts

7 Minimal # of Different Parts

8 Minimal # of Different Parts

9 Minimal # of Different Parts

10 Minimal # of Different Parts

11 Minimal # of Different Parts

12 Minimal # of Different Parts

13 Minimal # of Different Parts

14 Minimal # of Different Parts

15 15V 10A 35g 1Kg

16

17 Custom Electronics Power Remote

18 Custom Electronics Barometer

19 Off-the-Shelf Stuff Gyro

20 Off-the-Shelf Stuff Ultrasonic

21 Off-the-Shelf Stuff UWB RFID Laser

22 Off-the-Shelf Stuff Gumstix

23

24

25 Indoor Flight STARMAC Controller

26 Indoor Flight STARMAC Controller

27 Outdoor Flight Salzburg Controller

28 Outdoor Flight Salzburg Controller

29 More Recent: Yawing

30 More Recent: Yawing

31 Oops

32 Oops

33 Autonomous

34 Autonomous

35 A Cyber-Physical Server IP address location

36 A Cyber-Physical Server IP address location capabilities

37 A Cyber-Physical Server IP address location capabilities motion

38 A Cyber-Physical Server IP address location capabilities motion IP address location capabilities motion IP address location capabilities motion

39 A Cyber-Physical Server IP address location capabilities motion IP address location capabilities motion restricted IP address location capabilities motion

40 A Cyber-Physical Server IP address location capabilities motion IP address location capabilities motion restricted idealized IP address location capabilities motion

41 A Cyber-Physical Cloud [HotCloud 2010]

42 migration = flying A Cyber-Physical Cloud [HotCloud 2010]

43 Goals Multi-provider (10s): heterogeneous operations Multi-vehicle (100s): heterogeneous systems Multi-task (1000s): heterogeneous missions

44 High-Level Challenges Virtualization Infrastructure Salzburg Collaborative Control Berkeley Programming Language Berkeley, Salzburg

45 Virtualization Infrastructure ())*! " # $ %&' +

46 Virtualization Infrastructure Temporal Isolation Spatial Isolation Power Isolation ())* Migration Tracking! " # $ %&' +

47 There is a fundamental trade-off between quality and cost of time, space, power isolation

48 Time quality: response time jitter cost: scheduling overhead

49 Time quality: response time jitter cost: scheduling overhead Space quality: fragmentation jitter cost: management overhead

50 Time quality: response time jitter cost: scheduling overhead Space quality: fragmentation jitter cost: management overhead Power quality: power consumption jitter cost: total power consumption

51 Variable-Bandwidth Servers (VBS) Time [SIES09] VBS Overhead Accounting [RTAS10] Compact-fit Space [USENIX ATC08] Short-term Memory [Submitted11] Power-aware VBS Power [EMSOFT10] The Power of Isolation [Submitted11]

52 Time per-process lower/upper bounds on response times [SIES09] VBS Overhead Accounting [RTAS10] Compact-fit Space [USENIX ATC08] Short-term Memory [Submitted11] Power-aware VBS Power [EMSOFT10] The Power of Isolation [Submitted11]

53 Time per-process lower/upper bounds on response times [SIES09] account overhead in utilization and/or response times [RTAS10] Compact-fit Space [USENIX ATC08] Short-term Memory [Submitted11] Power-aware VBS Power [EMSOFT10] The Power of Isolation [Submitted11]

54 Time Space per-process lower/upper bounds on response times [SIES09] account overhead in utilization and/or response times [RTAS10] time- and space-predictable malloc and free [USENIX ATC08] Short-term Memory [Submitted11] Power-aware VBS Power [EMSOFT10] The Power of Isolation [Submitted11]

55 Time Space per-process lower/upper bounds on response times [SIES09] account overhead in utilization and/or response times [RTAS10] time- and space-predictable malloc and free [USENIX ATC08] refresh needed rather than deallocate not-needed objects [Submitted11] Power-aware VBS Power [EMSOFT10] The Power of Isolation [Submitted11]

56 Time Space Power per-process lower/upper bounds on response times [SIES09] account overhead in utilization and/or response times [RTAS10] time- and space-predictable malloc and free [USENIX ATC08] refresh needed rather than deallocate not-needed objects [Submitted11] save power while maintaining response times [EMSOFT10] The Power of Isolation [Submitted11]

57 Time Space Power per-process lower/upper bounds on response times [SIES09] account overhead in utilization and/or response times [RTAS10] time- and space-predictable malloc and free [USENIX ATC08] refresh needed rather than deallocate not-needed objects [Submitted11] save power while maintaining response times [EMSOFT10] per-process lower/upper bounds on power consumption [Submitted11]

58 The trade-off between isolation quality and cost extends to scalability on multicore systems

59 Multicore many concurrent data structures do not scale, e.g. a concurrent FIFO queue for scheduling

60 Multicore many concurrent data structures do not scale, e.g. a concurrent FIFO queue for scheduling linearizability makes a concurrent FIFO queue return the oldest element but limits scalability

61 Multicore many concurrent data structures do not scale, e.g. a concurrent FIFO queue for scheduling linearizability makes a concurrent FIFO queue return the oldest element but limits scalability k-linearizability allows it to scale but also return up to the k-oldest element

62 Multicore many concurrent data structures do not scale, e.g. a concurrent FIFO queue for scheduling linearizability makes a concurrent FIFO queue return the oldest element but limits scalability k-linearizability allows it to scale but also return up to the k-oldest element trades off semantics and scalability

63 Multicore many concurrent data structures do not scale, e.g. a concurrent FIFO queue for scheduling linearizability makes a concurrent FIFO queue return the oldest element but limits scalability k-linearizability allows it to scale but also return up to the k-oldest element trades off semantics and scalability can be done probabilistically

64 Multicore many concurrent data structures do not scale, e.g. a concurrent FIFO queue for scheduling linearizability makes a concurrent FIFO queue return the oldest element but limits scalability k-linearizability allows it to scale but also return up to the k-oldest element trades off semantics and scalability can be done probabilistically Non-linearizable Computing Breaks the Scalability Barrier [Submitted11]

65 Virtual Vehicle Demo by Florian Landolt and Andreas Rottmann

66 Virtual Vehicle Demo by Florian Landolt and Andreas Rottmann LAN

67 Virtual Vehicle Demo by Florian Landolt and Andreas Rottmann LAN Web Cam Web Cam

68 Laptop LAN Web Cam Web Cam

69 Laptop Multicast LAN Web Cam Web Cam

70 Laptop Migration LAN Web Cam Web Cam

71 Laptop LAN Web Cam Web Cam

72 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme)

73 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme) Inter-domain Multicast

74 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme) Inter-domain Multicast

75 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme) Inter-domain Multicast

76 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme) Inter-domain Multicast

77 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme) Inter-domain Multicast

78 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme) Inter-domain Multicast

79 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme) Inter-domain Multicast Domains register as receivers via XenStore

80 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme) Domains register as receivers via One XenSocket Inter-domain Multicast XenStore per receiver

81 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme) Inter-domain Multicast Domains register as receivers via XenStore One XenSocket per receiver Multicast is in user space (libxmc)

82 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme) Migration

83 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme) Migration VV scheme source

84 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme) VV scheme source Migration VV continuation

85 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme) VV scheme source Migration VV continuation network stack parameters

86 Virtual Vehicle Monitor (Xen, XenSockets, GUK, lwip, Newlib, Chibi Scheme) Migration VV scheme source VV continuation network stack parameters TCP connections (not used here)

87 3 VVs on 2 Servers

88 3 VVs on 2 Servers

89 Future Work (Simulating Motion, Flying Real Vehicles) LAN

90 Future Work (Simulating Motion, Flying Real Vehicles) LAN GPS Simulator Gyro

91 Future Work (Simulating Motion, Flying Real Vehicles) WLAN

92 Future Work (Simulating Motion, Flying Real Vehicles) WLAN read-only, scheduled real vehicle flight plans:

93 Future Work (Simulating Motion, Flying Real Vehicles) WLAN read-only, scheduled real vehicle flight plans: virtual-to-real vehicle allocation problem

94 Future Work (Simulating Motion, Flying Real Vehicles) WLAN read-only, scheduled real vehicle flight plans: virtual-to-real vehicle allocation problem read-write, on demand real vehicle flight plans:

95 Future Work (Simulating Motion, Flying Real Vehicles) WLAN read-only, scheduled real vehicle flight plans: virtual-to-real vehicle allocation problem read-write, on demand real vehicle flight plans: real-to-virtual vehicle allocation problem

96 Thank you Check out: eurosys2011.cs.uni-salzburg.at