Those numbers were for common use in Python tools and reflected my anecdotal experience at the time with normal Python tools. I'm sure that there are mechanisms to achieve faster speeds than what I experienced. That being said, here is a small example. In [1]: import multiprocessing In [2]: data = b'0' * 100000000 # 100 MB In [3]: from toolz import identity In [4]: pool = multiprocessing.Pool() In [5]: %time _ = pool.apply_async(identity, (data,)).get() CPU times: user 76 ms, sys: 64 ms, total: 140 ms Wall time: 252 ms This is about 400MB/s for a roundtrip On Thu, Sep 7, 2017 at 9:00 PM, Stephan Hoyer <shoyer@gmail.com> wrote:
On Thu, Sep 7, 2017 at 5:15 PM Nathaniel Smith <njs@pobox.com> wrote:
On Thu, Sep 7, 2017 at 4:23 PM, Nick Coghlan <ncoghlan@gmail.com> wrote:
The gist of the idea is that with subinterpreters, your starting point is multiprocessing-style isolation (i.e. you have to use pickle to transfer data between subinterpreters), but you're actually running in a shared-memory threading context from the operating system's perspective, so you don't need to rely on mmap to share memory over a non-streaming interface.
The challenge is that streaming bytes between processes is actually really fast -- you don't really need mmap for that. (Maybe this was important for X11 back in the 1980s, but a lot has changed since then :-).) And if you want to use pickle and multiprocessing to send, say, a single big numpy array between processes, that's also really fast, because it's basically just a few memcpy's. The slow case is passing complicated objects between processes, and it's slow because pickle has to walk the object graph to serialize it, and walking the object graph is slow. Copying object graphs between subinterpreters has the same problem.
This doesn't match up with my (somewhat limited) experience. For example, in this table of bandwidth estimates from Matthew Rocklin (CCed), IPC is about 10x slower than a memory copy: http://matthewrocklin.com/blog/work/2015/12/29/data-bandwidth
This makes a considerable difference when building a system do to parallel data analytics in Python (e.g., on NumPy arrays), which is exactly what Matthew has been working on for the past few years.
I'm sure there are other ways to avoid this expensive IPC without using sub-interpreters, e.g., by using a tool like Plasma ( http://arrow.apache.org/blog/2017/08/08/plasma-in-memory-object-store/). But I'm skeptical of your assessment that the current multiprocessing approach is fast enough.