会议6_OFI 集成共享内存卸载.pdf

《会议6_OFI 集成共享内存卸载.pdf》由会员分享，可在线阅读，更多相关《会议6_OFI 集成共享内存卸载.pdf（36页珍藏版）》请在三个皮匠报告上搜索。

1、OFI INTEGRATED SHARED MEMORY OFFLOAD2024 OFA Virtual WorkshopAlexia Ingerson,IntelShi Jin,AWSAmir Shehata,ORNLOVERVIEWPeer provider architecture provides a way for sharing resources between two or more providersTarget use case is for integrated shm offloadAWS efa provider was using shm provider deep

2、 within efa protocols to offload local communication but moved to peer provider infrastructureORNL has been developing a new“link”provider(LINKx)to allow any provider to offload to shm without having to manage two providersWhat is the peer provider and how does it work?What has changed since last ye

3、ar?How did AWS use it in its efa provider?What issues did they have and how did they solve them?How did using shm as a peer help efa?What is the link provider?What does a provider need in order to leverage LINKx support?What is the current status and direction of the provider?What are future extensi

4、ons?2 OpenFabrics AlliancePEER PROVIDER OVERVIEWExpose one endpoint to app while using two providersOne for external,internode communication(verbs,tcp,efa,cxi,etc)One for internal,intranode communication(shm)Share provider resourcesWrite to same CQUpdate same countersGet receive buffers from the sam

5、e receive context(SRX)Share addressing(e.g.fi_addr)All sharing and coordination is done internally,no application changes necessary“Owner”vs“peer”Owner owns resource and exports it for use by a peerPeer cannot directly access owner resource has to use imported ops3 OpenFabrics AlliancePEER PROVIDER

6、EXAMPLES4 OpenFabrics AllianceEFASHMSHM accesses EFA-owned resourcesEFA owns CQ,cntr,and SRXLINKxSHMCXILINKx owns CQ,cntr,and SRXSHM and CXI access LINKx resourcesimportimportimportEFA redirects intranode transfers to shmintranodeinternodeintranodeLINKx picks provider based on target addressLink pro