GPGPU-FedCloud
Status of accelerated computing in Clouds
Need efforts for additional development/support at all levels
- Chipset : HW virtualization support (otherwise some limitation)
- OS level: correct kernel configuration for the accelerators
- Hypervisor: configuration pass-through, vGPU
- CMFs: VM start, scheduler
- FedCloud facilities: accounting, information discovery
- Application: VM images with correct drivers for specific chipsets
Accelerators
GPGPU (General-Purpose computing on Graphical Processing Units)
NVIDIA GPU/Tesla/GRID, AMD Radeon/FirePro, Intel HD Graphics,...
Virtualization using VGA pass-through, vGPU (GPU partitioning) - NVIDIA GRID accelerators
- Shared Virtual GPUs (vGPU) http://www.nvidia.com/object/virtual-gpus.html
Intel Many Integrated Core Architecture
Xeon Phi Coprocessor
Virtualization using PCI pass-through
Specialized PCIe cards with accelerators
DSP (Digital Signal Processors)
FPGA (Field Programmable Gate Array)
Not commonly used in cloud environment
Hypervisors
QEMU/KVM
Supports only pass-through virtualization model
vGPU support is under development
Instructions for configuring passthrough in KVM (link ???)
Citrix XenServer 6, VMware ESXi 5.1
Support both pass-through and vGPU virtualization models
Limitations:
- vGPU support require certified server HW
- Live VM migration is not supported
- VM snapshot with memory is not supported
Security issues
- Non-standard PCI device functionality may render pass-through insecure (http://xenbits.xen.org/xsa/advisory-124.html)
Cloud Management Frameworks
Some work done with PCI passthrough
- OpenStack PCI passthrough (https://wiki.openstack.org/wiki/Pci_passthrough) meetings (https://wiki.openstack.org/wiki/Meetings/Passthrough)
vGPU is in very early stage
- Design document for GPU and vGPU support for CloudStack Guest VMs (https://cwiki.apache.org/confluence/display/CLOUDSTACK/GPU+and+vGPU+support+for+CloudStack+Guest+VMs)
- Request to support vGPU in OpenNebula (http://dev.opennebula.org/issues/3028)
Work to be done:
- Define VM types/flavors with attributes for GPGPU
- Modify VM start to allow passthrough or allocate vGPU
- Modify scheduler to allocate VMs with GPGPU correctly
VM images
VM images should contain proper drivers and libraries for specific accelerators
- Not transferable from site to site
More suitable approach is to use vanilla images with GPU support provided by cloud provider
- Using VM contextualization like cloud-init for installing applications
Or using VM snapshots
- May require support from site admins
FedCloud facilities
AppDB
- VM images are rather site-specific: any sense to use AppDB ?
Information discovery
- Should use similar GLUE2 scheme like grid sites with GPGPU
Accounting
- How to account GPU? (again to coordinate with grid)
Brokering, monitoring, VM management
Possible configuration
Dedicated cloud site with GPGPU
Homogenous: identical working nodes
Single VM type, single VM per node
- Simple configuration, no conflicting resources, no need to modify scheduler
Cloud site with OS level hypervisor
VMs can have direct access to hardware resources and share them
Limitation to the same OS/kernel
Related work
- GPU Passthrough Performance: A Comparison of KVM, Xen, VMWare ESXi, and LXC for CUDA and OpenCL Applications. http://www.isi.edu/sites/default/files/users/jwalters/papers/Cloud_2014.pdf
Progress
- May 2015
- Review of available technologies
- GPGPU virtualisation in KVM/QEMU
- Performance testing of passthrough
HW configuration: IBM dx360 M4 server with two NVIDIA Tesla K20 accelerators. Ubuntu 14.04.2 LTS with KVM/QEMU, PCI passthrough virtualization of GPU cards.
Tested application: NAMD molecular dynamics simulation (CUDA version), STMV test example (http://www.ks.uiuc.edu/Research/namd/).
Performance results: Tested application runs 2-3% slower in virtual machine compared to direct run on tested server. If hyperthreading is enabled on compute server, vCPUs have to be pinned to real cores so that whole cores will be dedicated to one VM. To avoid potential performance problems, hyperthreading should be switched off.
- June 2015
- Creating cloud site with GPGPU support
Configuration: master node, 2 worker nodes (IBM dx360 M4 servers, see above) Base OS: Ubuntu 14.04.2 LTS Hypervisor: KVM Middleware: Openstack Kilo
- July+August 2015
- Creating cloud site with GPGPU support
Cloud site created at keystone3.ui.savba.sk, master + two worker nodes, configuration reported above Creating VM images for GPGPU (based on Ubuntu 14.04, GPU driver and libraries)
- Testing cloud site with GPGPU support
Performance testing and tuning with GPGPU in Openstack - comparison performance of cloud-based VM with non-cloud virtualization and physical machine, - setting CPU flavor in Openstack nova (performance optimization) - Adjusting Openstack scheduler Starting process of integration of the site to EGI FedCloud - Keystone VOMS support being integrated - OCCI planned in September
- Next steps
Continue integration to EGI-FedCloud