Tessellation: Space-Time Partitioning in a Manycore Client OS

Transcription

1 Tessellation: Space-Time ing in a Manycore Client OS Rose Liu 1,2, Kevin Klues 1, Sarah Bird 1, Steven Hofmeyr 3, Krste Asanovic 1, John Kubiatowicz 1 1 Parallel Computing Laboratory, UC Berkeley 2 Data Domain 3 Lawrence Berkeley National Laboratory

2 Client Device Single-user device Runs a heterogeneous mix of interactive, real-time and batch applications simultaneously Generally battery constrained

3 Why a new Client OS? Enter the Manycore world Must address parallelism Current client OSs weren t designed for parallel applications Existing OSs addressing parallelism targets servers or HPC contexts, not clients Servers emphasis on throughput vs. Client emphasis on user experience/responsiveness HPC machine dedicated to one parallel application vs. Client runs many heterogeneous parallel applications Client - Longer battery life

4 Outline Why a new OS for Manycore Clients? A Case for Space-time ing Define space-time partitioning Use cases for space-time partitioning Implementing Space-Time ing in a Manycore OS Status

5 Spatial s Radio Isolated unit containing a subset of physical machine resources 5

6 Spatial s Radio Isolated unit containing a subset of physical machine resources 6

7 Spatial s Radio Isolated unit containing a subset of physical machine resources 7

8 Spatial s QoS enforced share of interconnect bandwidth Radio Isolated unit containing a subset of physical machine resources 8

9 Spatial s Energy or Power Budget Radio Isolated unit containing a subset of physical machine resources 9

10 Machine divided into spatial partitions Wireless radio 10

11 Put applications in spatial partitions Media Player Radio Browser 11

12 Benefits of spatial partitions Each app can run a custom user-level runtime for best performance Provides apps with resource guarantees for performance predictability Functional & Performance Isolation Natural unit for fault Media Player containment, energy management Radio Browser 12

13 Put OS Services in spatial partitions Media Player Network Driver Wireless radio Browser Filesystem 13

14 Put sub-components in spatial partitions Media Player Video decoder GUI Network Driver Wireless radio Browser Filesystem 14

15 Put virtual machines in spatial partitions Media Player Video decoder GUI Network Driver Wireless radio Filesystem Browser Windows VM 15

16 s need to communicate Spatial Communication occurs without a context switch Media Player Network Driver Wireless radio Video decoder GUI Browser Windows VM Filesystem 16

17 Communication Challenges Communication relaxes the isolation boundaries of partitions and introduces issues like: Security Service-level QoS and/or resource accounting of requestors within service partitions Media Player Video decoder GUI Network Driver Wireless radio Browser Windows VM Filesystem 17

18 Space-time partitioning virtualizes spatial partitions Context Switch Cost ~ Process Context Switch Cost Time multiplex at a coarse granularity to allow for user-level scheduling Media Player Video decoder Browser GUI Network Driver Windows VM Wireless radio Filesystem De-scheduled s 18

19 Space-Time Scheduling Descheduled s: s are dynamically resized while running without a reboot or application restart Time resources Real-time app put in low power state is always scheduled

20 Space-Time Scheduling Descheduled s: Challenges: 1. How to determine the right resource allocation for a partition? 2. What granularity to time multiplex each partition? Don t need to use same time quanta for all partitions. Time s are dynamically resized while running without a reboot or application restart 3. We can deschedule partitions from each type of resource independently. E.g. time multiplex off cores more frequently than multiplex partition data off caches. How to determine best policy?

21 Communication in space and time Media Player Video decoder GUI Network Driver Wireless radio Filesystem Browser Windows VM De-scheduled s 21

22 Outline Why a new OS for Manycore Clients? A Case for Space-time ing Implementing Space-Time ing in a Manycore OS (Tessellation) Status

23 Tessellation OS Media Player Video decoder GUI Network Driver Video Driver Wireless radio Browser Windows VM Filesystem Virus Checker De-scheduled s USB Driver 23

24 Tessellation Kernel Library OS Functionality Application Or OS Service Custom Scheduler Interconnect Bandwidth Message Passing Cache Physical CPUs Hardware ing Mechanisms Performance Counters 24

25 Tessellation Kernel Library OS Functionality Application Or OS Service Custom Scheduler Marshalls syscalls into messages for the respective OS Service App-specific scheduler for best parallel performance. (See Lithe talk on user-level scheduling.) Interconnect Bandwidth Message Passing Cache Physical CPUs Hardware ing Mechanisms Performance Counters 25

26 Tessellation Kernel Library OS Functionality Application Or OS Service Custom Scheduler Tessellation Kernel Interconnect Bandwidth Message Passing Cache Physical CPUs Hardware ing Mechanisms Performance Counters 26

27 Tessellation Kernel Library OS Functionality Application Or OS Service Custom Scheduler Management Layer Mechanism Layer (Trusted) Tessellation Kernel Interconnect Bandwidth Message Passing Cache Physical CPUs Hardware ing Mechanisms Performance Counters 27

28 Mechanism Layer Library OS Functionality Application Or OS Service Custom Scheduler Management Layer Mechanism Layer (Trusted) Configure Resources enforced by HW at runtime Tessellation Kernel Interconnect Bandwidth Message Passing Cache Physical CPUs Hardware ing Mechanisms Performance Counters 28

29 Mechanism Layer Library OS Functionality Application Or OS Service Custom Scheduler Management Layer Mechanism Layer (Trusted) Configure Resources enforced by HW at runtime Configure HW-supported Communication Tessellation Kernel Interconnect Bandwidth Message Passing Cache Physical CPUs Hardware ing Mechanisms Performance Counters 29

30 Management Layer Library OS Functionality Application Or OS Service Custom Scheduler Management Layer Mechanism Layer (Trusted) Configure Resources enforced by HW at runtime Allocator Configure HW-supported Communication Tessellation Kernel Interconnect Bandwidth Message Passing Cache Physical CPUs Hardware ing Mechanisms Performance Counters 30

31 Management Layer Library OS Functionality Application Or OS Service Custom Scheduler Resizing Callback API Management Layer Mechanism Layer (Trusted) Configure Resources enforced by HW at runtime Allocator Configure HW-supported Communication Tessellation Kernel Interconnect Bandwidth Message Passing Cache Physical CPUs Hardware ing Mechanisms Performance Counters 31

32 Management Layer Library OS Functionality Application Or OS Service Custom Scheduler Res. Reqs. Resizing Callback API Management Layer Mechanism Layer (Trusted) Configure Resources enforced by HW at runtime Allocator Configure HW-supported Communication Tessellation Kernel Interconnect Bandwidth Message Passing Cache Physical CPUs Hardware ing Mechanisms Performance Counters 32

33 Management Layer Library OS Functionality Application Or OS Service Comm Reqs Custom Scheduler Res. Reqs. Resizing Callback API Management Layer Mechanism Layer (Trusted) Scheduler Configure Resources enforced by HW at runtime Allocator Configure HW-supported Communication Tessellation Kernel Interconnect Bandwidth Message Passing Cache Physical CPUs Hardware ing Mechanisms Performance Counters 33

34 Management Layer Library OS Functionality Application Or OS Service Sched Reqs. Comm. Reqs Custom Scheduler Res. Reqs. Resizing Callback API Management Layer Mechanism Layer (Trusted) Scheduler Configure Resources enforced by HW at runtime Allocator Configure HW-supported Communication Tessellation Kernel Interconnect Bandwidth Message Passing Cache Physical CPUs Hardware ing Mechanisms Performance Counters 34

35 Outline Why a new OS for Manycore Clients? A Case for Space-time ing Define space-time partitioning Use cases for space-time partitioning Implementing Space-Time ing in a Manycore OS Status

36 Implementation status Basics of Tessellation kernel and primitive OS service up and running Provides rudimentary partition interface Boots on standard x86 hardware No I/O yet statically linked applications and kernel TOS OS Services User App 0 User App 1 Tesselation Kernel Manycore HW

37 Next Steps Build fast cross-core communication mechanisms for system calls Context-switch free system calls APIC driven message notification with shared memory Add support for the 19 newlib system calls in TOS OS Service partition TOS OS Services User App 0 Newlib C User App 1 Newlib C Syscall Syscall

38 Intermediate Infrastructure TOS OS Service doesn t have all drivers, so run BSD with existing drivers on one core to service I/O from TOS OS Service Tessellation runs on rest of the cores Pinned Shared Channel Existing I/O Drivers I/O Server Process BSD Kernel TOS OS Service User App 0 Newlib C Syscall Manycore HW Syscall User App 1 Newlib C

39 Acknowledgements We would like to thank the entire ParLab OS group and several people at LBNL for their invaluable contribution to the ideas presented here. Research supported by Microsoft Award # and Intel Award # and by matching funding from U.C. Discovery (Award #DIG ). This work has also been in part supported by an NSF Graduate Research Fellowship.

40 Thanks! Questions? Media Player Video decoder GUI Network Driver Wireless radio Filesystem Browser Windows VM De-scheduled s 40