M 3 Microkernel-based System for Heterogeneous Manycores

Transcription

1 M 3 Microkernel-based System for Heterogeneous Manycores Nils Asmussen MKC, 06/29/ / 35

2 Outline 1 Introduction 2 Data Transfer Unit 3 Prototype Platforms 4 M 3 5 Summary 2 / 35

3 Outline 1 Introduction 2 Data Transfer Unit 3 Prototype Platforms 4 M 3 5 Summary 3 / 35

4 Motivation Intel Knights Landing 4 / 35

5 Motivation Intel Knights Landing Qualcom Snapdragon / 35

6 What s the Problem? DSP ARM big Audio Decoder Intel Xeon Intel Xeon FPGA DSP ARM LITTLE 5 / 35

7 What s the Problem? DSP ARM big Audio Decoder Intel Xeon Kernel Intel Xeon Kernel FPGA DSP ARM LITTLE 5 / 35

8 What s the Problem? DSP ARM big Kernel Audio Decoder Intel Xeon Kernel Intel Xeon Kernel FPGA DSP ARM LITTLE Kernel 5 / 35

9 What s the Problem? ARM big Kernel Intel Xeon Kernel Intel Xeon Kernel ARM LITTLE Kernel 5 / 35

10 The Goal Treat all compute units (CU) as first-class citizens: 1 Run untrusted code without causing harm 2 Access operating system services 3 Interact as the master with other CUs 6 / 35

11 First-class Citizenchip as Enabler Pipe communication between arbitrary CUs Use parallism on GPUs for FS operations Direct access to accelerators from the net... 7 / 35

12 My Approach DSP ARM big Audio Decoder Intel Xeon Intel Xeon FPGA DSP ARM LITTLE 8 / 35

13 My Approach DSP ARM big Audio Decoder Intel Xeon Intel Xeon FPGA DSP ARM LITTLE 8 / 35

14 My Approach DSP ARM big Audio Decoder Intel Xeon Intel Xeon FPGA DSP ARM LITTLE 8 / 35

15 My Approach PE PE PE PE DSP ARM big Audio Decoder Intel Xeon PE PE PE PE Intel Xeon FPGA DSP ARM LITTLE 8 / 35

16 My Approach PE PE PE PE DSP App ARM Kernel big Audio App Decoder Intel App Xeon PE PE PE PE Intel App Xeon FPGA App DSP App ARM App LITTLE 8 / 35

17 My Approach PE PE PE PE DSP App ARM Kernel big Audio App Decoder Intel App Xeon PE PE PE PE Intel App Xeon FPGA App DSP App ARM App LITTLE 8 / 35

18 Outline 1 Introduction 2 Data Transfer Unit 3 Prototype Platforms 4 M 3 5 Summary 9 / 35

19 Data Transfer Unit Supports memory access and message passing Provides a number of endpoints Each endpoint can be configured for: 1 Accessing a piece of memory (contiguous range, byte granular) 2 Receiving messages into a ringbuffer 3 Sending messages to a receiving endpoint Configuration only by kernel, usage by application Direct reply on received messages Credits are used to prevent DoS attacks 10 / 35

20 Data Transfer Unit Example DSP App ARM big Kernel S R 11 / 35

21 Data Transfer Unit Example DSP App ARM big Kernel S R 11 / 35

22 Data Transfer Unit Example DSP App ARM big Kernel S R 11 / 35

23 Outline 1 Introduction 2 Data Transfer Unit 3 Prototype Platforms 4 M 3 5 Summary 12 / 35

24 Introduction Data Transfer Unit Prototype Platforms M3 Summary Tomahawk 2 and 4 PE Xtensa LX4 PE R Data SPM PE R PE PE R Instr. SPM PE PE R R R R Ctrl. PE R R DRAM Cores attached to NoC with No privileged mode No MMU, no caches, but SPM T2: simple ; T4: most features 13 / 35

25 Linux M 3 runs on Linux using it as a virtual machine A process simulates a PE, having two threads (CPU + ) s communicate over UNIX domain sockets No accuracy because Programs are directly executed on host Data transfers have huge overhead compared to HW Very useful for debugging and early prototyping 14 / 35

26 gem5 Modular platform for computer-system architecture research Supports various ISAs (x86, ARM, Alpha, SPARC,... ) Cycle-accurate simulation Has an out-of-order core model We built a for gem5 Support for caches and virtual memory We also added hardware accelerators 15 / 35

27 gem5 Example Configuration x86 L1$ PE x86 PE DRAM ME L2$ IO$ L1$ IO$ Accel PE Accel PE VM SPM L1$ 16 / 35

28 Outline 1 Introduction 2 Data Transfer Unit 3 Prototype Platforms 4 M 3 5 Summary 17 / 35

29 Introduction M 3 uses ideas from L4, but implemented from scratch Provides mechanisms for PEs, memory and communication Drivers, filesystems,... are implemented as services Kernel manages permissions, using capabilities enforces permissions (communication, memory access) Kernel has no ISA-specific code 18 / 35

30 Capabilities M 3 has the following capabilities: Send: send messages to a receive EP Receive: receive messages from send EPs ory: access remote memory via Mapping: access remote memory via load/store Service: create sessions Session: exchange caps with service VPE: execute code on a PE 19 / 35

31 Capability Exchange Kernel provides syscalls to create, exchange and revoke caps There are two ways to exchange caps: 1 Directly with another VPE (typically, a child VPE) 2 Over a session with a service The kernel offers two operations: 1 Delegate: send capability to somebody else 2 Obtain: receive capability from somebody else 20 / 35

32 Communication Overview Kernel: PE0 Core VPE1: PE1 Recv Cap VPE2: PE2 Send Cap Receiver: PE1 Core Recv Gate Sender: PE2 Core Send Gate cmdreg cmdreg adds cmdreg EP EP buffer header data EP credits unread occup channel target label configuration of endpoints to establish a channel 21 / 35

33 Communication System call to create send cap for specific EP and VPE Typically created for own VPE Binds a label to the send cap for secure identification Can then be delegated to someone else To use a send cap, an EP needs to be configured The kernel provides a syscall for that purpose This indirection allows EP multiplexing and blocking 22 / 35

34 ory Syscall reqmem can be used to request remote memory Like with send caps, an EP needs to be configured to access it Afterwards, the can be instructed to read/write Behaves like RDMA (no SW involved) Syscall derivemem to create a subset of mem cap 23 / 35

35 VPEs Syscall createvpe to create a new VPE The kernel assigns the VPE to a suitable PE, if possible App gets VPE cap and memory cap for the whole PE memory Syscall to start it and wait for its completion (exit) Can exchange capabilities with the VPE before starting it: e.g., delegate inputs after it s finished: e.g., obtain results 24 / 35

36 VPEs Examples Executing ELF-Binaries VPE vpe (" test "); char * args [] = {"/ bin / hello ", " foo ", " bar "}; vpe. exec (3, args ); 25 / 35

37 VPEs Examples Executing ELF-Binaries VPE vpe (" test "); char * args [] = {"/ bin / hello ", " foo ", " bar "}; vpe. exec (3, args ); Asynchronous Lambdas VPE vpe (" test "); Gate mem = Gate :: create_ global (0 x1000, RW); vpe. delegate ( CapRngDesc ( mem. sel (), 1)); vpe. run_async ([& mem ]() { mem. read (buf, sizeof ( buf )); cout << " Done reading!\ n"; }); 25 / 35

38 Filesystem: m3fs Filesystem service is implemented outside of kernel m3fs is (currently) an in-memory filesystem PE1 PE2 PE3 m3fs App kernel FS 26 / 35

39 Filesystem: m3fs Filesystem service is implemented outside of kernel m3fs is (currently) an in-memory filesystem PE1 PE2 PE3 open m3fs App kernel FS 26 / 35

40 Filesystem: m3fs Filesystem service is implemented outside of kernel m3fs is (currently) an in-memory filesystem PE1 PE2 PE3 obtain m3fs App kernel FS 26 / 35

41 Filesystem: m3fs Filesystem service is implemented outside of kernel m3fs is (currently) an in-memory filesystem obtain PE1 PE2 PE3 obtain m3fs App kernel FS 26 / 35

42 Filesystem: m3fs Filesystem service is implemented outside of kernel m3fs is (currently) an in-memory filesystem PE1 PE2 PE3 m3fs App kernel FS read/write 26 / 35

43 Pipes: Direct Pipe writer reader 27 / 35

44 Pipes: Direct Pipe Shared ory writer reader msg passing 28 / 35

45 Pipes: Indirect Pipe writer reader writer reader reader 29 / 35

46 Pipes: Indirect Pipe Shared ory writer reader msg passing service 30 / 35

47 Files on Top of Capabilities Example libm3 provides abstractions to work with files, pipes,... Every VPE has a mountspace and file descriptors These can be inherited to child VPEs Pure convenience: security is based on capabilities VPE vpe ("my vpe "); vpe. fds () ->set ( STDIN_FD, pipe_fd ); vpe. fds () ->set ( STDOUT_FD, VPE :: self ().fds () ->get ( STDOUT_FD )); vpe. obtain_fds (); vpe. run_async ([] { File *in = VPE :: self ().fds () ->get ( STDIN_FD ); //... }); 31 / 35

48 Demo Demo 32 / 35

49 Outline 1 Introduction 2 Data Transfer Unit 3 Prototype Platforms 4 M 3 5 Summary 33 / 35

50 Ongoing Work Multiple instances of the kernel/services (by Matthias Hille) Port of the musl C library (by Sherif Shehabaldin) Integration of a NIC (by Georg Kotheimer) Integration of an IDE controller (by Lukas Landgraf) 34 / 35

51 Summary +M 3 is a new point in the design space for systems Supports untrusted code and OS services on all PEs Supports direct communication between all PEs abstracts heterogeneity and provides common interface M 3 kernel controls s remotely M 3 is available at 35 / 35