Peter Andreas Entschev

You're right @madsbk . This is essentially being triggered by having https://github.com/rapidsai/dask-cuda/blob/ec80f97b2f01328df6512c9cc1379a784e13f0b9/dask_cuda/tests/test_proxy.py#L27-L35 in the context of the test file, commenting out the `ProxifyHostFile` instantiation also prevents this from happening. IOW,...

Also pinging @ayushdg @jnke2016 @randerzander who may have other feature requests in mind.

> FYI: I'm looking into the related problem of visualizing GPU spilling. You mean you want to visualize it but there's no way to do that, or there's a problem...

I've never really used `--default-stream per-thread`, what's the behavior if you happen to access the same data on multiple threads, do you need explicit synchronization? If so, how would that...

It seems like `--default-stream per-thread` could then be helpful if we assign multiple threads per worker (currently we only use one in dask-cuda). My original idea was to have the...

I began with some testing and changes to enable PTDS support in https://github.com/rapidsai/rmm/pull/633 and https://github.com/cupy/cupy/pull/4322 .

@charlesbluca @beckernick @VibhuJawa @ayushdg @randerzander just so you're aware of this, in case this shows up somehow in any workflows where you may test TCP performance. I believe this should...

> Curious how was it determined compression was happening with GPU data? Asking as this is [explicitly disabled in Distributed](https://github.com/dask/distributed/blob/3551d1574c9cd72d60197cc84dd75702ebcfec54/distributed/protocol/cuda.py#L28) and has been this way for a while. Would be...

> TCP _currently_ requires host to device copying regardless of whether there is compression or not. So disabling compression wouldn't fix that. Yes, I know, and this is not just...

Particularly I find frustrating those defaults that are hard to really know beforehand, the compression default itself is a great example, something changed in Dask (pulling lz4 by default) that...