Enable RISC V capability in cloud computing



all right our next speaker is G paying Wong from Huawei he is the operations open source operations manager and he's going to speak to us about risk v and cloud computing and while he's getting set up real quick I just want to point out we've got the dinner that I mentioned earlier for those that weren't here and then the dinner that we're having tonight is not strictly for risk five Foundation members that's open to everyone so you know we want to make sure everybody comes to that dinner so and as mentioned we'll have escorts directing us after the events this afternoon so without further ado okay I know I'm the last guy standing between you and lunch so I promise I'll be quick and also there's this will be a high-level non-technical probably the weirdest talk you will experience in the workshop so I'm leaving a open-sourcing in Far Away mostly we are doing open source cloud computing infrastructure so this talk will be about how to enable like risk five days accelerators in cloud computing so we will not cover like risk 5 CPUs or in IOT scenario ok so the most specific accelerators is definitely becoming a in a global trend and also cloud computing is everywhere basically you now you offer anything through account so we are seeing like AWS at one instance offers FPGA instance or a juror has the catapult infrastructure which heavily utilized abbreviate to power the crowd and also open source in front I'm so now basically you can build a open source cloud infrastructure from ground up if you have like capable developers so everything looks sunny it's actually not so during our development we found that there is a like big problem that for hardware developers the the concept of software ecosystem rarely reach beyond the device SDK and application developers barely like care about the hardware and you leave cloud infrastructure developers especially open source of core infrastructure developers in the middle that have no clue how to support all these type of new things so just give you a sense we just had the biggest ever cucum barcelona you can see the number accelerator is definitely not on the radar at the moment so that really bothers me so there are a lot of open source rigs five base accelerators for example the fire team project is a awesome project you can this is a site actually a copy from their talk this year about like provide video a through the fire scene also you have for example open a sorority so I think what is 5 the the most important thing is that with a open source instruction set you can actually build a pipeline that is best fit for your application for example if you want to try new things like graph neural network probably you will be able to use risk 5 to write a awesome ASIC for that so how do we support risk 5 based accelerators in open source cloud computing in OpenStack we are leading developing a protocol cyborg which is providing a general manager framework for all kinds of accelerators FPGA GPU you name it and also in Cuba Nettie's we are also brainstorming with Intel and NVIDIA to also have a similar project that can provide this general management framework and from development one thing is very interesting for us is that the most thing we cared about for the schedulers of the exhibition is actually the metadata and all of the metadata you see on the right hand side the most important thing is to prodigy so like GPU topology or @pg topology this is very important like the information we need to get from the accelerators to make the scheduler to to schedule the workload to the most appropriate node so basically we really want to make reso five basic sir sorry sir a first-class citizen and so in a nutshell we need however experts like you guys the help work with us and defining like what kind of capabilities a risk 5 core based accelerator can report to the cloud observation platform you'll be amazed how much information that the cloud orchestration can actually utilize for example in Cuba nineties like topology I just talked about inaudible Intel developers also talked about like socket closeness also Finity power there is like unlimited thing that we can extract from the accelerator and for the cloud observation platform so this is actually came out from my my conversation with a rhombic developers at the open happy poof in the OCP summit back in March so they demonstrate there's a big red cable and we talked about how basically we want to expose as little as possible for the application so that the application don't view the underlying change however the latter point is rarely mentioned is that for cloud you actually need to provide as much information as possible because we need to schedule the workload to the most appropriate node so this year and also starting a really lean open source like dialog because I found I find that I need to like talk to different people and bring them together actually to help with a acceleration support so this idea I actually also discuss with Lisa I think when she was three days into the job we will touch upon all the related foundations but the efforts itself is pretty lightweight because we have all the Foundation's so to give you an example for example for edge computing you have a use case that you can use communities to utilize the Linux boot or boot for like automated provisioning of resources onto a OCP oai compliant accelerator cluster with the probably core five course that written unit five instruction set so this like example tells you like how much we can leverage the open source project to build a full stack reference implementation so if we have like similar use case or just interesting ideas you can submit so we will formalize that as a formula and through the discussion of formula we'll identify if there's a gap in the upstream community we'll go to the upstream community do the work if not there is a process called HDI P similar to the PP or PID process that probably we can trying to standardize some of the things that we find is edible so send his system summat I will be there if you can drive a car there so I think we can have like a lot a lot of one-on-one time and also we have the first co-located event during the OSS China actually for that we experiments with like using github issue for safety it's pretty awesome I think the open transparency and basically pure okay so feel free to reach out to me and we really need our expert to help green risk five basic servitors as a first-class citizen for the cloud infrastructure and yep thank you all right do we have any questions

Be First to Comment

Leave a Reply

Your email address will not be published. Required fields are marked *