Skip to content

Latest commit

 

History

History
371 lines (279 loc) · 13.5 KB

README.md

File metadata and controls

371 lines (279 loc) · 13.5 KB
Raw

Automatic Virtualization of Accelerators

AvA is a research system for virtualizing general API-controlled accelerators automatically, developed in SCEA lab at University of Texas at Austin. AvA is prototyped on KVM and QEMU, compromising compatibility with automation for classic API remoting systems, and introducing hypervisor interposition for resource management and strong isolation.

This repository is the main codebase of AvA. We host customized Linux kernel, QEMU, LLVM, and sets of benchmarks in separate repositories.

Related repositories

Setup code tree

Clone the repositories into a single directory:

git clone [email protected]:utcs-scea/ava.git
cd ava
git clone https://github.com/utcs-scea/ava-benchmarks benchmark
git clone https://github.com/utcs-scea/ava-llvm llvm
git clone https://github.com/utcs-scea/ava-qemu qemu
git clone https://github.com/utcs-scea/ava-kvm  kvm

Install

Dependencies

The following packages are required to build and run all AvA components and benchmarks:

  • Ubuntu (apt install): git libssl-dev libglib2.0-dev libpixman-1-dev opencl-headers curl bc build-essential libclang-7-dev clang-7 ctags caffe-cpu pssh python3 python3-pip virtualenv indent
  • Python3 (pip3 install): blessings toposort astor numpy(==1.15.0)

AvA was fully tested on Ubuntu 18.04 with GCC 5.5+, Python 3.6+, Bazel 0.26.1, and customized LLVM 7.0, Linux kernel 4.14, and QEMU 3.1. The system also works in Ubuntu 16.04 with additional care of Python 3.6 installation for CAvA scripts.

Hardware and API

The following hardware and APIs are virtualized with AvA (excluding manually implemented Python forwarding):

API framework Hardware
OpenCL 1.2 NVIDIA GTX 1080 / AMD RX 580
CUDA 10.0 (driver) NVIDIA GTX 1080
CUDA 10.0 (runtime) NVIDIA GTX 1080
TensorFlow 1.12 C Intel Xeon E5-2643
TensorFlow 1.14 Python NVIDIA GTX 1080
NCSDK v2 Intel Movidius NCS v1 & v2
GTI SDK 4.4.0.3 Gyrfalcon 2803 Plai Plug
QuickAssist 1.7 Intel QuickAssist
Custom FPGA on AmorphOS AWS F1

Environment

Export the following variables in bash profile:

Name Example Explanation
AVA_ROOT /project/ava Path to AvA source tree
AVA_CHANNEL SHM Transport channel (SHM|TCP|VSOCK|LOCAL)
AVA_MANAGER_ADDR 0.0.0.0:3333 (guestlib only) AvA manager's address
AVA_WPOOL TRUE Enable API server pool
DATA_DIR /project/rodinia/data Path to Rodinia dataset

More environment variables are introduces in CAvA and benchmarks' documentations.

Install Kernel (KVM)

The host kernel needs to be replaced to enable VSOCK interposition and resource management (LOCAL channel can run with native kernel without interposition). We provide kernel configuration and Kbuild script in $AVA_ROOT/kbuild. KVM and VHOST_VSOCK modules are required.

## In $AVA_ROOT/kbuild
cp 4.14.config .config && make -f Makefile.setup
make -j16 && make modules
sudo make modules_install ; sudo make install

Reboot and verify the installed kernel by uname -r. Because AvA depends on VHOST_VSOCK module, it needs to be loaded by sudo modprobe vhost_vsock before starting any AvA components. Or it can be automatically loaded by

## In /etc/modules-load.d/modules.conf
vhost_vsock
## Reboot the machine
sudo reboot

Compile QEMU

Compile QEMU in the normal way as below. The VM boot script will point to $AVA_ROOT/qemu directory, so there is no need to install the compiled binary.

## In $AVA_ROOT/qemu
./configure
make -j16

Compile LLVM

Build the artifacts into $AVA_ROOT/llvm/build which are required by CAvA. You do not need to install them. CAvA will find them in their build locations.

## In $AVA_ROOT/llvm
mkdir build
cd build
cmake -DLLVM_ENABLE_PROJECTS=clang \
        -G "Unix Makefiles" ../llvm
make

Compile generated stack and API server

CAvA is written in Python3. To verify the Python package installation and CAvA's help menu, please run ./nwcc -v in $AVA_ROOT/cava.

## In $AVA_ROOT/cava
./nwcc samples/opencl.nw.c
./nwcc samples/cuda.nw.c \
         -I /usr/local/cuda-10.0/include
./nwcc samples/cudart.nw.c \
         -I headers
         -I /usr/local/cuda-10.0/include \
         `pkg-config --cflags glib-2.0`
./nwcc samples/mvnc.nw.c

The guest libraries and workers for those APIs are generated into cl_nw, cu_nw, cudart_nw, mvnc_nw, correspondingly. They are compiled by make R=1 for release version, or just make with debugging messages enabled.

API server manager

Simply execute make in $AVA_ROOT/worker.

Support for Ubuntu 16.04

Installing the missing packages will just work fine:

wget -O - https://apt.llvm.org/llvm-snapshot.gpg.key|sudo apt-key add -
sudo add-apt-repository \"deb http://apt.llvm.org/xenial/ llvm-toolchain-xenial-7 main\"
sudo add-apt-repository ppa:jonathonf/python-3.6
sudo apt update
sudo apt install libclang-7-dev
sudo apt install python3.6

SOL_NETLINK socket option is supported in the 16.04 kernel, but it is not in the libc headers. However libc passes this option though to the kernel so defining it here should be just as good.

Zero-copy host driver (optional)

Simply execute make in $AVA_ROOT/hostdrv. The zero copy mode is enabled if the compiled driver (zcopy.ko) is installed before the VM is booted.

Benchmarking

This section will show an example of utilizing virtualized accelerators in VM.

Start API server manager

Link the compiled API server binary to $AVA_ROOT/worker and start manager.

## In $AVA_ROOT/worker
ln -s $AVA_ROOT/cava/cl_nw/worker worker
sudo -E ./manager

Setup virtual machine

The example boot scripts are located in $AVA_ROOT/vm. Update DIR_QEMU in scripts/environment for compiled QEMU binary, and IMAGE_FILE for the VM disk image.

The IP address is specified by tapno in scripts/options. By default, the first VM can be accessed by ssh localhost -p 2222.

Before boot the VM by ./run.sh, make sure that sudo -E ./manager has been started in $AVA_ROOT/worker, and vhost_vsock kernel module has been installed (by sudo modprobe vhost_vsock). In the QEMU parameter list, you need to

  • Enable KVM: -enable-kvm -machine accel=kvm -cpu host,kvm=on
  • Assign guest-cid (starts from 3): -device vhost-vsock-pci,guest-cid=5
  • Enable AvA transport device: -device ava_vdev

Setup guest environment

Once the guest VM is booted, the guestdrv source and compiled guestlib need to be copied to VM. Furthermore, the benchmark must load guestlib instead of original framework library at runtime. Run $AVA_ROOT/vm/setup.sh to setup guestdrv and guestlib in guest automatically:

## In $AVA_ROOT/vm
## Copy necessary files (e.g., $AVA_ROOT/cava/cl_nw) to
## VM ($GUEST_AVA_ROOT defined in the script), compile
## and install guestdrv
./setup.sh cl_nw

Then export AVA_ROOT and DATA_DIR in guest. Make sure AVA_CHANNEL and AVA_WPOOL are consistent between guest and host.

Note: It is necessary to install essential tool chains to compile benchmarks in guest, e.g. CUDA nvcc.

In the prototype, because the virtual transport device is built on top of vhost_vsock, AvA requires the VSOCK version in the guest matches that in the host. Because AvA host uses 4.14 kernel, the guest VM needs to use kernel 4.10+.

Compile benchmark and link against libguestlib.so

It is highly recommended to run the benchmark first in the host with native library to verify the framework installation.

Example for running a single benchmark:

## In (guest) $AVA_ROOT/benchmark/rodinia/opencl/backprop
make && ./run

To run the same benchmark in host, comment \lstinline|make.mk:7-8| before compiling the benchmark:

## In $AVA_ROOT/benchmark/rodinia/opencl/util/make.mk
# OPENCL_LIB = $(AVA_ROOT)/cava/cl_nw
# LIB_NAME = guestlib

Minimal remoting system

AvA supports a minimal remoting system without replacing the host driver or running a VM. The minimal system does not have hypervisor interposition and both guestlib and worker run in the host.

To enable the minimal system:

  • Switch the communication channel to LOCAL mode: export AVA_CHANNEL=LOCAL.
  • Run sudo -E ./manager_tcp instead of ./manager in $AVA_ROOT/worker.
  • Run the application (compiled against libguestlib.so) in host.

Hard-coded parameters

AvA moved a part of configurations to be environment variables, while kept others which are not commonly used hard-coded.

Most configurations can be found in $AVA_ROOT/devconf.h, where the parameter names and comments imply the meanings. For the demo, you can use the default settings.

To enable AvA system-wide debugging prints, define DEBUG in $AVA_ROOT/include/debug.h.

AvA requires a clean rebuild after changing the hard-coded configurations.

Environment variables

The following variables must be set and consistent in the guest VM and host.

  • Communication channel export AVA_CHANNEL=LOCAL|SHM|VSOCK|TCP. The default value is LOCAL when AVA_CHANNEL is unset. The parameter is read by both the guestlib and the worker to determine the transport method. To use LOCAL or TCP channel, the manager_tcp (instead of manager) is required to be started.

  • AvA manager host address export AVA_MANAGER_ADDR=<Server name or IP addreses:port>. The variable must be set for guestlib. If the address is barely a port, the server name will use localhost.

  • Worker pooling export AVA_WPOOL=TRUE|FALSE. The default is FALSE when AVA_WPOOL is unset. This applies to the manager to pre-initialize a worker pool in the host.

The following variables must be set for special features in only the host.

  • Migration random call id export AVA_MIGRATION_CALL_ID=r<limit> (such as r100) or <number>. When set to r<limit> the runtime will migrate at a random call with number less than limit. When set to <number> the runtime will migrate at exactly call number.

Configuring CAvA's Generated Code

AvA API specification can contain C flags to enable or disable some features during prototyping, for example, ava_cflags(-DAVA_RECORD_REPLAY) enables the record-and-replay feature.

  • AVA_RECORD_REPLAY: when defined, API record-and-replay is enabled.

  • AVA_API_FUNCTION_CALL_RESOURCE: when defined, resource reporting annotations take effects.

  • AVA_DISABLE_HANDLE_TRANSLATION: when defined, the handle id translation is disabled and the actual address is used as the handle id.

  • AVA_BENCHMARKING_MIGRATE: when defined, an invocation will be selected to start the migration benchmarking. Set environment variable AVA_MIGRATION_CALL_ID to r<limit> (such as r100) to choose a random call, and to <count> to choose a specified call.

  • AVA_PRINT_TIMESTAMP: when defined, APIs with ava_time_me will print the timestamps for the invocation's forwarding stages. The current sequence of an invocation is:

    Guestlib: before_marshal,
Guestlib: before_send_command,
Worker: after_receive_command,
Worker: after_unmarshal,
Worker: after_execution,
Worker: after_marshal,
Guestlib: after_receive_result.

The makefile generated by CAvA accepts a "release build" flag used as follows: make RELEASE=1 (or for the very lazy make R=1). When set this flag will enable optimization and disable debugging. In release mode the build will:

  • use -O2.
  • disable -g.
  • disable debug printing.
  • disable assertions and many error check that should only happen if there is a bug in AvA.

This mode will almost certainly be faster (since it will allow the compiler to aggressively optimize the primary data path). However, it will make debugging basically impossible, since crashes will likely be due to data corruption instead of a nice assert. The main reason for disabling -g is to remind us to rebuild in debug mode before trying to debug.

Publications

Yu, Hangchen, Arthur M. Peters, Amogh Akshintala, and Christopher J. Rossbach. "Automatic Virtualization of Accelerators." In Proceedings of the Workshop on Hot Topics in Operating Systems, pp. 58-65. ACM, 2019.

[Todo: add conference version]

Developers and contributors

Name Affiliation Role Contact
Hangchen Yu UT Austin Main developer [email protected]
Arthur M. Peters UT Austin Main developer [email protected]
Amogh Akshintala UNC
Tyler Hunt UT Austin
Christopher J. Rossbach UT Austin & VMware Research Advisor [email protected]