www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - Thread-Multiplexing for D

reply "Bienlein" <jeti789 web.de> writes:
Hello,

I once had a look at Go and D and can't see a reason to choose Go
except for language built-in CSP and thread-multiplexing. But
thread-multiplexing in Go is really a killer feature, especially
when developing server-side applications that need to scale with
the number of requests (from the Internet). Several startups
chose therefore Go. It could have been D.

So my point is that it would be good to have thread multiplexing
in D as well. It exists for Java, see
http://hawtdispatch.fusesource.org/. And then there is the
original from Apple:
http://en.wikipedia.org/wiki/Grand_Central_Dispatch.

I think server-side applications that need to scale and are not
for enterprise computing (aka company internal Java clusters) is
a new kind of application for which people could also decide to
choose D for just like Go.

Putting in some thread-multiplexing into D is not easy and I
won't have the time for doing that. But I would like to suggest
to start an effort in that direction or think it over whether
that would be something to have.

Regards, Bienlein
Sep 12 2013
next sibling parent reply =?UTF-8?B?U8O2bmtlIEx1ZHdpZw==?= <sludwig outerproduct.org> writes:
Am 12.09.2013 16:12, schrieb Bienlein:
 Hello,

 I once had a look at Go and D and can't see a reason to choose Go
 except for language built-in CSP and thread-multiplexing. But
 thread-multiplexing in Go is really a killer feature, especially
 when developing server-side applications that need to scale with
 the number of requests (from the Internet). Several startups
 chose therefore Go. It could have been D.

 So my point is that it would be good to have thread multiplexing
 in D as well. It exists for Java, see
 http://hawtdispatch.fusesource.org/. And then there is the
 original from Apple:
 http://en.wikipedia.org/wiki/Grand_Central_Dispatch.

 I think server-side applications that need to scale and are not
 for enterprise computing (aka company internal Java clusters) is
 a new kind of application for which people could also decide to
 choose D for just like Go.

 Putting in some thread-multiplexing into D is not easy and I
 won't have the time for doing that. But I would like to suggest
 to start an effort in that direction or think it over whether
 that would be something to have.

 Regards, Bienlein

There is an existing implementation already: http://vibed.org/ I didn't hear the term thread-multiplexing before, though. Maybe that should be added somewhere to make it more searchable.
Sep 12 2013
next sibling parent reply =?ISO-8859-1?Q?S=F6nke_Ludwig?= <sludwig outerproduct.org> writes:
Am 12.09.2013 19:55, schrieb Sean Kelly:
 On Sep 12, 2013, at 8:34 AM, Bienlein <jeti789 web.de> wrote:

 About thread-multiplexing... You find a lot in Google when
 searching for socket multplexing, but not when searching for
 thread-multiplexing. Maybe I coined the term myself (don't know
 any more) when reading the section here:
 http://golang.org/doc/effective_go.html#goroutines

 "Goroutines are multiplexed onto multiple OS threads so if one
 should block, such as while waiting for I/O, others continue to
 run. Their design hides many of the complexities of thread
 creation and management."

The trick in D is that because statics are thread-local by default, any multiplexed app like this that expects its static data to remain consistent across calls is likely to fail. I've mentioned fiber-local storage here in the past, but it's a tricky problem. But I think it's one that we will need to sort out for things like this to work as the user expects them to.

Until we have something built-in, I've implemented a template based solution for task local storage (same as fiber local storage, but the lifetime is different as fibers get recycled for consecutive tasks): http://vibed.org/api/vibe.core.core/TaskLocal
Sep 13 2013
parent reply Dmitry Olshansky <dmitry.olsh gmail.com> writes:
13-Sep-2013 13:57, Sönke Ludwig пишет:
 Am 12.09.2013 19:55, schrieb Sean Kelly:
 On Sep 12, 2013, at 8:34 AM, Bienlein <jeti789 web.de> wrote:

 About thread-multiplexing... You find a lot in Google when
 searching for socket multplexing, but not when searching for
 thread-multiplexing. Maybe I coined the term myself (don't know
 any more) when reading the section here:
 http://golang.org/doc/effective_go.html#goroutines

 "Goroutines are multiplexed onto multiple OS threads so if one
 should block, such as while waiting for I/O, others continue to
 run. Their design hides many of the complexities of thread
 creation and management."

The trick in D is that because statics are thread-local by default, any multiplexed app like this that expects its static data to remain consistent across calls is likely to fail. I've mentioned fiber-local storage here in the past, but it's a tricky problem. But I think it's one that we will need to sort out for things like this to work as the user expects them to.

Until we have something built-in, I've implemented a template based solution for task local storage (same as fiber local storage, but the lifetime is different as fibers get recycled for consecutive tasks): http://vibed.org/api/vibe.core.core/TaskLocal

I do hope O(n) is a typo and it should be O(1). Anyhow what n stands here for? -- Dmitry Olshansky
Sep 13 2013
next sibling parent =?UTF-8?B?U8O2bmtlIEx1ZHdpZw==?= <sludwig outerproduct.org> writes:
Am 13.09.2013 12:57, schrieb Dmitry Olshansky:
 13-Sep-2013 13:57, Sönke Ludwig пишет:
 Am 12.09.2013 19:55, schrieb Sean Kelly:
 On Sep 12, 2013, at 8:34 AM, Bienlein <jeti789 web.de> wrote:

 About thread-multiplexing... You find a lot in Google when
 searching for socket multplexing, but not when searching for
 thread-multiplexing. Maybe I coined the term myself (don't know
 any more) when reading the section here:
 http://golang.org/doc/effective_go.html#goroutines

 "Goroutines are multiplexed onto multiple OS threads so if one
 should block, such as while waiting for I/O, others continue to
 run. Their design hides many of the complexities of thread
 creation and management."

The trick in D is that because statics are thread-local by default, any multiplexed app like this that expects its static data to remain consistent across calls is likely to fail. I've mentioned fiber-local storage here in the past, but it's a tricky problem. But I think it's one that we will need to sort out for things like this to work as the user expects them to.

Until we have something built-in, I've implemented a template based solution for task local storage (same as fiber local storage, but the lifetime is different as fibers get recycled for consecutive tasks): http://vibed.org/api/vibe.core.core/TaskLocal

I do hope O(n) is a typo and it should be O(1). Anyhow what n stands here for?

Indeed, stupid typo.
Sep 13 2013
prev sibling parent =?UTF-8?B?U8O2bmtlIEx1ZHdpZw==?= <sludwig outerproduct.org> writes:
Am 13.09.2013 12:57, schrieb Dmitry Olshansky:
 13-Sep-2013 13:57, Sönke Ludwig пишет:
 Am 12.09.2013 19:55, schrieb Sean Kelly:
 On Sep 12, 2013, at 8:34 AM, Bienlein <jeti789 web.de> wrote:

 About thread-multiplexing... You find a lot in Google when
 searching for socket multplexing, but not when searching for
 thread-multiplexing. Maybe I coined the term myself (don't know
 any more) when reading the section here:
 http://golang.org/doc/effective_go.html#goroutines

 "Goroutines are multiplexed onto multiple OS threads so if one
 should block, such as while waiting for I/O, others continue to
 run. Their design hides many of the complexities of thread
 creation and management."

The trick in D is that because statics are thread-local by default, any multiplexed app like this that expects its static data to remain consistent across calls is likely to fail. I've mentioned fiber-local storage here in the past, but it's a tricky problem. But I think it's one that we will need to sort out for things like this to work as the user expects them to.

Until we have something built-in, I've implemented a template based solution for task local storage (same as fiber local storage, but the lifetime is different as fibers get recycled for consecutive tasks): http://vibed.org/api/vibe.core.core/TaskLocal

I do hope O(n) is a typo and it should be O(1). Anyhow what n stands here for?

n would be the number of variables declared. There is a linear initialization cost and in the naive predecessor implementation, an AA with Variants was used to store the values, which is why I've mentioned efficiency at all.
Sep 13 2013
prev sibling parent =?UTF-8?B?U8O2bmtlIEx1ZHdpZw==?= <sludwig outerproduct.org> writes:
Am 12.09.2013 16:57, schrieb Bienlein:
 Ah! Cool :-). They are saying that they are using libevent, see
 http://vibed.org/features#performance. I see... Although only
 shitty software comes from my country such as SAP, this is a
 company that seems to develop some cool stuff. So I hope that D
 will also pop up in job ads over here some day ;-).

 -- Bienlein

I too hope that this will happen one day, but I fear that the company is still in its infancy :) * BTW there is at least one company (sociomantic) in Germany with an office in Berlin that uses D. * disclaimer: I might be affiliated with vibe.d and that company ;)
Sep 13 2013
prev sibling next sibling parent "Bienlein" <jeti789 web.de> writes:
Ah! Cool :-). They are saying that they are using libevent, see
http://vibed.org/features#performance. I see... Although only
shitty software comes from my country such as SAP, this is a
company that seems to develop some cool stuff. So I hope that D
will also pop up in job ads over here some day ;-).

-- Bienlein
Sep 12 2013
prev sibling next sibling parent "Bienlein" <jeti789 web.de> writes:
About thread-multiplexing... You find a lot in Google when
searching for socket multplexing, but not when searching for
thread-multiplexing. Maybe I coined the term myself (don't know
any more) when reading the section here:
http://golang.org/doc/effective_go.html#goroutines

"Goroutines are multiplexed onto multiple OS threads so if one
should block, such as while waiting for I/O, others continue to
run. Their design hides many of the complexities of thread
creation and management."

-- Bienlein
Sep 12 2013
prev sibling next sibling parent Sean Kelly <sean invisibleduck.org> writes:
On Sep 12, 2013, at 8:34 AM, Bienlein <jeti789 web.de> wrote:

 About thread-multiplexing... You find a lot in Google when
 searching for socket multplexing, but not when searching for
 thread-multiplexing. Maybe I coined the term myself (don't know
 any more) when reading the section here:
 http://golang.org/doc/effective_go.html#goroutines
=20
 "Goroutines are multiplexed onto multiple OS threads so if one
 should block, such as while waiting for I/O, others continue to
 run. Their design hides many of the complexities of thread
 creation and management."

The trick in D is that because statics are thread-local by default, any = multiplexed app like this that expects its static data to remain = consistent across calls is likely to fail. I've mentioned fiber-local = storage here in the past, but it's a tricky problem. But I think it's = one that we will need to sort out for things like this to work as the = user expects them to.=
Sep 12 2013
prev sibling next sibling parent "Dicebot" <public dicebot.lv> writes:
On Friday, 13 September 2013 at 11:06:35 UTC, Sönke Ludwig wrote:
 BTW there is at least one company (sociomantic) in Germany with 
 an office in Berlin that uses D.

And they are hiring ;)
Sep 13 2013
prev sibling next sibling parent "Bienlein" <jeti789 web.de> writes:
The "thread-multiplexing" in Go is described here:

https://docs.google.com/document/d/1TTj4T2JO42uD5ID9e89oa0sLKhJYD0Y_kqxDv3I3XMw/edit

The sources are here:

http://code.google.com/p/go/source/browse/src/pkg/runtime/proc.c?r=01acf1dbe91f673f6308248b8f45ec0564b1d751

It should be possible to takes this approach from Go and bring it
to D. Just an idea ...
Oct 09 2013
prev sibling next sibling parent "qznc" <qznc web.de> writes:
On Wednesday, 9 October 2013 at 13:36:41 UTC, Bienlein wrote:
 The "thread-multiplexing" in Go is described here:

 https://docs.google.com/document/d/1TTj4T2JO42uD5ID9e89oa0sLKhJYD0Y_kqxDv3I3XMw/edit

 The sources are here:

 http://code.google.com/p/go/source/browse/src/pkg/runtime/proc.c?r=01acf1dbe91f673f6308248b8f45ec0564b1d751

 It should be possible to takes this approach from Go and bring 
 it
 to D. Just an idea ...

Is this different to task parallelism as implemented in std.parallelism [0]? The downside of D is that the rest of system is not integrated. For example, if you do a blocking syscall then the threadpool is not increased to compensate for the blocked thread. [0] http://dlang.org/phobos/std_parallelism.html
Oct 09 2013
prev sibling parent "Bienlein" <jeti789 web.de> writes:
On Wednesday, 9 October 2013 at 15:47:11 UTC, qznc wrote:
 On Wednesday, 9 October 2013 at 13:36:41 UTC, Bienlein wrote:
 The "thread-multiplexing" in Go is described here:

 https://docs.google.com/document/d/1TTj4T2JO42uD5ID9e89oa0sLKhJYD0Y_kqxDv3I3XMw/edit

 The sources are here:

 http://code.google.com/p/go/source/browse/src/pkg/runtime/proc.c?r=01acf1dbe91f673f6308248b8f45ec0564b1d751

 It should be possible to takes this approach from Go and bring 
 it
 to D. Just an idea ...

Is this different to task parallelism as implemented in std.parallelism [0]? The downside of D is that the rest of system is not integrated. For example, if you do a blocking syscall then the threadpool is not increased to compensate for the blocked thread. [0] http://dlang.org/phobos/std_parallelism.html

Yes, right. This is the big difference between CSP-style channels in Go compared to libdispatch in Apple's C/C++.
Oct 09 2013