Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

TSC BoF: Accelerators Discussion - SFO17-320

113 views

Published on

Session ID: SFO17-320
Session Name: TSC BoF: Accelerators Discussion - SFO17-320
Speaker:
Track:


★ Session Summary ★
Discussion on the various hardware acceleration efforts in Linaro. What are we doing now and how should this evolve?
---------------------------------------------------
★ Resources ★
Event Page: http://connect.linaro.org/resource/sfo17/sfo17-320/
Presentation:
Video:
---------------------------------------------------

★ Event Details ★
Linaro Connect San Francisco 2017 (SFO17)
25-29 September 2017
Hyatt Regency San Francisco Airport

---------------------------------------------------
Keyword:
'http://www.linaro.org'
'http://connect.linaro.org'
---------------------------------------------------
Follow us on Social Media
https://www.facebook.com/LinaroOrg
https://twitter.com/linaroorg
https://www.youtube.com/user/linaroorg?sub_confirmation=1
https://www.linkedin.com/company/1026961

Published in: Technology
  • Very interesting!.... STARTUPS...Send your pitchdeck to thousands of VC's and Angel's with just 1 click. Visit: Angelvisioninvestors.com
       Reply 
    Are you sure you want to  Yes  No
    Your message goes here
  • Be the first to like this

TSC BoF: Accelerators Discussion - SFO17-320

  1. 1. linaro Committee Confidential @ 2017 SFO17-320 TSC Accelerators Discussion David A Rusling Francois, Martin
  2. 2. ENGINEERS AND DEVICES WORKING TOGETHER linaro Committee Confidential @ 2017 Overview ● Current hardware acceleration work in Linaro (and elsewhere) ○ LNG ○ LEG ● Discussion ○ Are any activities common (or should any activities be common?) ○ Anything missing?
  3. 3. ENGINEERS AND DEVICES WORKING TOGETHER linaro Committee Confidential @ 2017 LNG perspective ● Use cases ○ Cloud RAN: matrix compression on FPGA ○ Mobile Edge computing: Video acceleration ○ CPE (enterprise or consumer): IPsec acceleration ● Solutions needed ○ Linux kernel ■ complex to add because of crash risks and hence upstream delays (if accepted) ■ Required latencies are such that “one size fits all” kernel approach do not meet performance ○ Userland ■ Diverse ways to match performance goals and bug containment (as opposed to kernel crash) ■ Not a Linux first class citizen, security issues because not using IOMMUs ● Vision ○ Stop kernel bypass, start kernel driver cooperation: secure userland direct dataplane IO ○ Linux kernel: always in charge of IO controlplane ○ Movement started ■ vfio_mdev for virtio-gpu ■ Mellanox libiverbs DPDK PMD on bifurcated drivers ■ LNG: vf-netdev for ODP, DPDK, Netmap, Zephyr...
  4. 4. ENGINEERS AND DEVICES WORKING TOGETHER linaro Committee Confidential @ 2017 LNG activities ● Mimic OpenVPN: IPsec acceleration with tap interface ○ Transparent to Linux and applications ○ Huge efficiency benefits (1 A53 core + IPsec HW to deal with 20Gbps IMIX) ■ Can handle fat pipes ■ 5x less expensive per secured Gbps that ANY other solution ○ Plan to be delivered as a LEG ERP 2018.1 package and other forms ■ Strongswan? OpenVPN for the IPsec ESP tunnel? ENEA Linux? ● net_mdev ○ Leverage VFIO to expose queues (and only queues) to userland ○ Prototyping, checking relationship with other technologies (WrapDrive, dma_buf, Netm ● virtio-fpga ○ Implement ETSI NFV IFA002 specifications ■ Standard load of bitstream through virtio-fpga -> obtain passthrough device on the fly ■ Leverage that device directly ■ May be create classes of passthrough devices (network…) ○ Current status: positive industry contacts, public support from Cisco, Ericsson, Nokia ○ Not limited to LNG very broad use cases
  5. 5. ENGINEERS AND DEVICES WORKING TOGETHER linaro Committee Confidential @ 2017 Coherent acceleration ● Technologies: Arm: CCIX, Intel: PPC: CAPI/CCIX ● Radical change of IO model, even if it does look like it ● If drivers don’t evolve (metadata for IO…) most of benefits will be lost ● New model is about cacheline movement control which means in-depth understanding of cache coherency protocol by developers ● peer to peer connectivity to chain accelerators and IOMMU relationship is very unknown at this stage ○ PCIe Access Control Services (ACS) was created to solve this on PCIe
  6. 6. ENGINEERS AND DEVICES WORKING TOGETHER linaro Committee Confidential @ 2017 LMG Accelerator Interest ● Mobile Tensorflow ○ https://www.tensorflow.org/mobile/ ○ “makes sense when there is a poor or missing network connection, or where sending continuous data to a server would be too expensive” ● Use case ○ Image processing/identification is a common use case ○ Voice input ● Vulkan, OpenCL-next ○ Vulkan is strategic (part of AOSP) ○ OpenCL hasn’t seen adoption by handset vendors ○ Is the merging of Vulkan & OpenCL-next what the ARM ecosystem needs
  7. 7. linaro Committee Confidential @ 2017 Thank You #SFO17 BUD17 keynotes and videos on: connect.linaro.org For further information: www.linaro.org
  8. 8. ENGINEERS AND DEVICES WORKING TOGETHER linaro Committee Confidential @ 2017 Alternative Content Slide ● If you have a more succinct message, use this slide
  9. 9. ENGINEERS AND DEVICES WORKING TOGETHER linaro Committee Confidential @ 2017 This Slide is for when Two Columns are Needed ● You can use this for two columns of bullets ● Or you can replace one column with an image or diagram ● This could be a second column of bullets ● Or it could be a table, image or graphic
  10. 10. ENGINEERS AND DEVICES WORKING TOGETHER linaro Committee Confidential @ 2017 Use this Slide for a Large Image
  11. 11. ENGINEERS AND DEVICES WORKING TOGETHER linaro Committee Confidential @ 2017 Useful Logos Download Hi Res logos from here* to use on your slides *http://link.linaro.org/logos
  12. 12. ENGINEERS AND DEVICES WORKING TOGETHER linaro Committee Confidential @ 2017 Some Images ● Download Hi Res photos from here* to use on your slides *http://link.linaro.org/groupphotos

×