[prev in list] [next in list] [prev in thread] [next in thread] 

List:       xml-dev
Subject:    Re: [xml-dev] Re: ***SPAM*** [xml-dev] Re: The Goals of XML at 25, and the one thing that XML now ne
From:       Rick Jelliffe <rjelliffe () allette ! com ! au>
Date:       2021-07-22 17:02:27
Message-ID: CADUdYQX0B3Pu5fwzJayv82pX0RgcTM4a3DsyvY0E-xKGAK7yLw () mail ! gmail ! com
[Download RAW message or body]

On Fri, 23 Jul. 2021, 01:41 Peter Flynn, <peter@silmaril.ie> wrote:

>
>.  It seems to
> > implicitly assume that the entirety of the source document is already
> > available,
>
> Which in the document field (eg Humanities encoded texts, books,
> articles in journals, etc) is almost certainly true.
>

SIMD parallel operations don't care whether the document is all there.

I think there is a connection with streaming in some cases of parallelism.

For example, lets say you have a huge document over a slow net, and your
processing scenario is to parallel lex the documents using an NVIDIA GPU,
which lets you run 20,000 threads at the same time, in synced groups of 32
threads and then use a conventional multi-core CPU to do the parse.

 But you do need to corordinate the running of the thread groups with the
availability of the data. If you wait till the last packet arrives before
starting the lexing or parsing, you may need a lot of good parallelism
before you can reduce latency.  (But if you are on an uncongested server
and your interest is highest throughput rather than lowest latency, the
parallel approaches can still useful inasmuchas they can harness all the
available cores. )

Rick

[Attachment #3 (text/html)]

<div dir="auto"><div><br><br><div class="gmail_quote"><div dir="ltr" \
class="gmail_attr">On Fri, 23 Jul. 2021, 01:41 Peter Flynn, &lt;<a \
href="mailto:peter@silmaril.ie">peter@silmaril.ie</a>&gt; wrote:<br></div><blockquote \
class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc \
solid;padding-left:1ex"></blockquote></div></div><div dir="auto"><br></div><div \
dir="auto"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 \
0 .8ex;border-left:1px #ccc solid;padding-left:1ex">&gt;.   It seems to<br> &gt; \
implicitly assume that the entirety of the source document is already<br> &gt; \
available,<br> <br>
Which in the document field (eg Humanities encoded texts, books,<br>
articles in journals, etc) is almost certainly true.<br></blockquote></div></div><div \
dir="auto"><br></div><div dir="auto">SIMD parallel operations don&#39;t care whether \
the document is all there.</div><div dir="auto"><br></div><div dir="auto">I think \
there is a connection with streaming in some cases of parallelism.  </div><div \
dir="auto"><br></div><div dir="auto">For example, lets say you have a huge document \
over a slow net, and your processing scenario is to parallel lex the documents using \
an NVIDIA GPU, which lets you run 20,000 threads at the same time, in synced groups \
of 32 threads and then use a conventional multi-core CPU to do the parse.    \
</div><div dir="auto"><br></div><div dir="auto">  But you do need to corordinate the \
running of the thread groups with the availability of the data. If you wait till the \
last packet arrives before starting the lexing or parsing, you may need a lot of good \
parallelism before you can reduce latency.   (But if you are on an uncongested server \
and your interest is highest throughput rather than lowest latency, the parallel \
approaches can still useful inasmuchas they can harness all the available cores. \
)</div><div dir="auto"><br></div><div dir="auto">Rick</div><div \
dir="auto"><br></div><div dir="auto"><div class="gmail_quote"><blockquote \
class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc \
solid;padding-left:1ex"> </blockquote></div></div></div>



[prev in list] [next in list] [prev in thread] [next in thread]