[libfabric-users] [chuck at ece.cmu.edu: utility provider breaks fi_wait()]
jianxin.xiong at intel.com
Tue Sep 29 13:19:21 PDT 2020
It's about the provider code in libfabric. For one part, the address vector access is not thread safe. There may be other aspects related to the active message mechanism used for implementing RMA and atomics, but I can't recall the details right now.
> -----Original Message-----
> From: Chuck Cranor <chuck at ece.cmu.edu>
> Sent: Tuesday, September 29, 2020 12:17 PM
> To: Xiong, Jianxin <jianxin.xiong at intel.com>
> Cc: Hefty, Sean <sean.hefty at intel.com>; libfabric-users at lists.openfabrics.org; Pritchard Jr., Howard <howardp at lanl.gov>; Jerome
> Soumagne <jsoumagne at hdfgroup.org>
> Subject: Re: [chuck at ece.cmu.edu: utility provider breaks fi_wait()]
> On Mon, Sep 28, 2020 at 10:24:01PM +0000, Xiong, Jianxin wrote:
> > FI_THREAD_SAFE is not guaranteed in the psm provider. In the past a
> > lot of effort have been put into the psm2 provider to make it thread
> > safe, but similar work was not done to the psm provider.
> just to be clear: are your main concern on threading just in the PSM provider code in libfabric, or are you also concerned with the
> libinfinipath and libpsm_infinipath libraries that are part of PSM itself?
More information about the Libfabric-users