| Author |
Message |
Seongbae Park
Guest
|
 Posted:
Wed Jul 13, 2005 8:26 pm Post subject:
A newly published paper on hardware scout thread. |
|
|
Sun's Rock architects recently published a paper on IEEE Micro:
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1463182
High-Performance Througput Computing
Chaudhry, S. Caprioli, P. Yip, S. Tremblay, M.
IEEE Micro
Publication Date: May-June 2005
Volume: 25, Issue: 3
On page(s): 32- 45
ISSN: 0272-1732
It mostly talks about the hardware scout thread.
Although it sometimes reads like a marketing document,
it has some technical content
that Sun observers (and others as well) in comp.arch may find interesting.
Personally, I'm a little surprised that they published this paper
at this point in time...
Anyway, this paper should at least give some second thought
to those who think Sun gave up on the single thread performance.
--
#pragma ident "Seongbae Park, compiler, http://blogs.sun.com/seongbae/" |
|
| Back to top |
|
 |
Rich Walker
Guest
|
| Posted:
Thu Jul 14, 2005 12:15 am Post subject:
Re: A newly published paper on hardware scout thread. |
|
|
Seongbae Park <Seongbae.Park@Sun.COM> writes:
| Quote: | Sun's Rock architects recently published a paper on IEEE Micro:
|
No they didn't.
They made a paper available to subscribers to a proprietary information
database.
Quote:
| Quote: |
You are not logged in.
Guests may access Abstract records free of charge.
You must log in to access:
\u2022 Advanced or Author Search
\u2022 CrossRef Search
\u2022 AbstractPlus Records
\u2022 Full Text PDF
\u2022 Full Text HTML
I appreciate that your USENET postings are regularly full of |
high-quality information, but in this case you appear to be advertising
a commercial service.
If you can find a reference to this material that we can read, then this
is probably a really interesting thread :->
cheers, Rich.
--
rich walker | Shadow Robot Company | rw@shadow.org.uk
technical director 251 Liverpool Road |
need a Hand? London N1 1LX | +UK 20 7700 2487
www.shadow.org.uk/products/newhand.shtml |
|
| Back to top |
|
|
Chris Colohan
Guest
|
| Posted:
Thu Jul 14, 2005 6:11 am Post subject:
Re: A newly published paper on hardware scout thread. |
|
|
Rich Walker <rw@shadow.org.uk> writes:
| Quote: | Seongbae Park <Seongbae.Park@Sun.COM> writes:
Sun's Rock architects recently published a paper on IEEE Micro:
No they didn't.
They made a paper available to subscribers to a proprietary information
database.
|
Huh? IEEE Micro is a paper journal. Go to your local university
library and look it up if you don't subscribe. Or, you can access it
online if you are an IEEE member (with online access) or are willing
to pay the per-document fee.
Publish != give away for free. How do you think publishers stay in
business?
Chris
--
Chris Colohan Email: chris@colohan.ca PGP: finger colohan@cs.cmu.edu
Web: www.colohan.com Phone: (412)268-4751 |
|
| Back to top |
|
|
Andrew Reilly
Guest
|
| Posted:
Thu Jul 14, 2005 6:20 am Post subject:
Re: A newly published paper on hardware scout thread. |
|
|
On Thu, 14 Jul 2005 00:40:18 +0100, Rich Walker wrote:
| Quote: | Seongbae Park <Seongbae.Park@Sun.COM> writes:
Sun's Rock architects recently published a paper on IEEE Micro:
No they didn't.
They made a paper available to subscribers to a proprietary information
database.
|
Or of the Journal in question (IEEE Micro). When did that stop being
"published"?
Most good University libraries will probably have access or a copy, if you
don't want to subscribe personally.
--
Andrew |
|
| Back to top |
|
|
Robert Myers
Guest
|
| Posted:
Thu Jul 14, 2005 7:58 am Post subject:
Re: A newly published paper on hardware scout thread. |
|
|
Seongbae Park wrote:
| Quote: | Sun's Rock architects recently published a paper on IEEE Micro:
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1463182
High-Performance Througput Computing
Chaudhry, S. Caprioli, P. Yip, S. Tremblay, M.
IEEE Micro
Publication Date: May-June 2005
Volume: 25, Issue: 3
On page(s): 32- 45
ISSN: 0272-1732
|
The "scout thread" is apparently not a separate cooperative thread, but
the main thread in run-ahead mode, with execution set to resume from a
checkpoint when the delinquent load that triggered the checkpoint is
fulfilled.
RM |
|
| Back to top |
|
|
Nick Maclaren
Guest
|
| Posted:
Thu Jul 14, 2005 8:15 am Post subject:
Re: A newly published paper on hardware scout thread. |
|
|
In article <1121309929.226865.177350@g49g2000cwa.googlegroups.com>,
Robert Myers <rbmyersusa@gmail.com> wrote:
| Quote: |
The "scout thread" is apparently not a separate cooperative thread, but
the main thread in run-ahead mode, with execution set to resume from a
checkpoint when the delinquent load that triggered the checkpoint is
fulfilled.
|
Please don't let POSIX rot your fine brain. Hoare's seminal paper
was about cooperative PROCESSES, and the concept to which he was
referring is one that includes POSIX threads. When people started
to talk about lightweight threads, they were referring to things
more like the Niagara/Rock scout threads than POSIX threads, which
are assuredly not lightweight.
Yes, there are two threads, in the sense that the flow of execution
is independent.
You are correct that there are not two POSIX threads, which I should
much prefer to be called sub-processes, as they are almost but not
entirely unlike lightweight threads.
Regards,
Nick Maclaren. |
|
| Back to top |
|
|
Patrick Schaaf
Guest
|
| Posted:
Thu Jul 14, 2005 8:15 am Post subject:
Re: A newly published paper on hardware scout thread. |
|
|
Chris Colohan <colohan+@cs.cmu.edu> writes:
| Quote: | Huh? IEEE Micro is a paper journal. Go to your local university
library and look it up if you don't subscribe. Or, you can access it
online if you are an IEEE member (with online access) or are willing
to pay the per-document fee.
|
As somebody with casual interest in those topics, working at a company
with almost no interest in those topics, I won't do any of that.
So SUN lost at least one reader for this paper, and thus failed
to gain a potential advocate for their potential innovation.
Of course, they don't know me. I'm insignificant. And we didn't buy
much SUN gear in the past, and probably won't do so in the future.
But they just lost a chance to change those probabilities slightly.
| Quote: | Publish != give away for free. How do you think publishers stay in
business?
|
Who cares about publishers staying in business, except for the
publishers themselves?
More to the point, what does SUN, as the author of the paper, gain by
supporting this publisher and excluding more casual readership?
best regards
Patrick |
|
| Back to top |
|
|
Robert Myers
Guest
|
| Posted:
Thu Jul 14, 2005 4:15 pm Post subject:
Re: A newly published paper on hardware scout thread. |
|
|
Nick Maclaren wrote:
| Quote: | In article <1121309929.226865.177350@g49g2000cwa.googlegroups.com>,
Robert Myers <rbmyersusa@gmail.com> wrote:
The "scout thread" is apparently not a separate cooperative thread, but
the main thread in run-ahead mode, with execution set to resume from a
checkpoint when the delinquent load that triggered the checkpoint is
fulfilled.
Please don't let POSIX rot your fine brain. Hoare's seminal paper
was about cooperative PROCESSES, and the concept to which he was
referring is one that includes POSIX threads. When people started
to talk about lightweight threads, they were referring to things
more like the Niagara/Rock scout threads than POSIX threads, which
are assuredly not lightweight.
Yes, there are two threads, in the sense that the flow of execution
is independent.
|
But the flow of execution is not independent. A single thread
encounters a load stall, check points, continues to execute in
run-ahead mode, returns to the checkpoint when the load completes, and
proceeds forward, making final results. Whatever separate threads
there may be as defined by context, they apparently execute
sequentially; and they are not, in that sense, independent.
RM |
|
| Back to top |
|
|
Nick Maclaren
Guest
|
| Posted:
Thu Jul 14, 2005 4:15 pm Post subject:
Re: A newly published paper on hardware scout thread. |
|
|
In article <1121340258.463699.51070@g14g2000cwa.googlegroups.com>,
"Robert Myers" <rbmyersusa@gmail.com> writes:
|> >
|> > Yes, there are two threads, in the sense that the flow of execution
|> > is independent.
|>
|> But the flow of execution is not independent. A single thread
|> encounters a load stall, check points, continues to execute in
|> run-ahead mode, returns to the checkpoint when the load completes, and
|> proceeds forward, making final results. Whatever separate threads
|> there may be as defined by context, they apparently execute
|> sequentially; and they are not, in that sense, independent.
Well, communicating sequential processes (sorry!) aren't independent,
either. What I meant was that they are independent except for their
specified interactions which, in this case, are memory accesses.
Let's use the term "asynchronously parallel" instead.
My point is that processes in the sense that Hoare used the term
are an abstract form of (an instantiation of a Von Neumann machine),
and the bracketted term is close enough to a Unix process as to make
little practical difference. The original lightweight thread models
were different, and some of them included asynchronously parallel
execution of dependent code fragments (which is what you have here).
POSIX threads started off as fairly lightweight, and accumulated
so much bloat that they have ended up being effectively processes,
but sharing most resources and some parameters. I have heard
reports that some people have invented the term "fibre" (perhaps
"fiber") to mean what the term thread was orginally used for.
However, in contexts other than POSIX, the term "thread" often
still means what it used to.
Regards,
Nick Maclaren. |
|
| Back to top |
|
|
Robert Myers
Guest
|
| Posted:
Thu Jul 14, 2005 4:15 pm Post subject:
Re: A newly published paper on hardware scout thread. |
|
|
Nick Maclaren wrote:
| Quote: |
My point is that processes in the sense that Hoare used the term
are an abstract form of (an instantiation of a Von Neumann machine),
and the bracketted term is close enough to a Unix process as to make
little practical difference. The original lightweight thread models
were different, and some of them included asynchronously parallel
execution of dependent code fragments (which is what you have here).
POSIX threads started off as fairly lightweight, and accumulated
so much bloat that they have ended up being effectively processes,
but sharing most resources and some parameters. I have heard
reports that some people have invented the term "fibre" (perhaps
"fiber") to mean what the term thread was orginally used for.
However, in contexts other than POSIX, the term "thread" often
still means what it used to.
|
Do you have any talent for constructing synoptic overhead charts? Or
maybe it's already been constructed: the taxonomy of the flora and
fauna of the thread zoo.
I think that what we've got here is a straightforward implementation of
run-ahead, with no clever complications of speculative slices and what
not (schemes which would result in simultaneous execution of more than
one instruction stream--whatever you want to call those streams).
If there is this much performance gain to be had, it's a little
puzzling as to why Sun is the first to implement the most obvious
scheme. Maybe Intel has been too busy trying to be clever?
RM |
|
| Back to top |
|
|
Nick Maclaren
Guest
|
| Posted:
Thu Jul 14, 2005 4:15 pm Post subject:
Re: A newly published paper on hardware scout thread. |
|
|
In article <1121344005.200945.254450@g49g2000cwa.googlegroups.com>,
"Robert Myers" <rbmyersusa@gmail.com> writes:
|>
|> Do you have any talent for constructing synoptic overhead charts? Or
|> maybe it's already been constructed: the taxonomy of the flora and
|> fauna of the thread zoo.
Given the amount of miscegenation involved, it would be a nightmare
task!
|> I think that what we've got here is a straightforward implementation of
|> run-ahead, with no clever complications of speculative slices and what
|> not (schemes which would result in simultaneous execution of more than
|> one instruction stream--whatever you want to call those streams).
I haven't seen that article, and I the talks I have been to were
rather woolly, but that could well be right. On the other hand,
I am not sure that it is.
|> If there is this much performance gain to be had, it's a little
|> puzzling as to why Sun is the first to implement the most obvious
|> scheme. Maybe Intel has been too busy trying to be clever?
Yes. I am unconvinced about the gains, and await evidence. I
suspect that it may fall at the same jump as the Itanic.
Regards,
Nick Maclaren. |
|
| Back to top |
|
|
Chris Colohan
Guest
|
| Posted:
Thu Jul 14, 2005 4:15 pm Post subject:
Re: A newly published paper on hardware scout thread. |
|
|
"Robert Myers" <rbmyersusa@gmail.com> writes:
| Quote: | If there is this much performance gain to be had, it's a little
puzzling as to why Sun is the first to implement the most obvious
scheme. Maybe Intel has been too busy trying to be clever?
|
I would guess that the performance gain is all relative to an in-order
CPU. As soon as you go to an out-of-order superscalar design the OOO
core finds other ways of hiding memory latency.
Also, Sun is claiming a performance advantage on _commercial
workloads_ (by which I presume they mean transaction processing).
From what I have seen, Intel has cared much more about SPEC-like
workloads and desktop PC workloads in the past --- I suspect that
Sun's hardware may not do as well on such workloads (and Sun may not
care).
Chris
--
Chris Colohan Email: chris@colohan.ca PGP: finger colohan@cs.cmu.edu
Web: www.colohan.com Phone: (412)268-4751 |
|
| Back to top |
|
|
Robert Myers
Guest
|
| Posted:
Thu Jul 14, 2005 4:15 pm Post subject:
Re: A newly published paper on hardware scout thread. |
|
|
Chris Colohan wrote:
| Quote: | "Robert Myers" <rbmyersusa@gmail.com> writes:
If there is this much performance gain to be had, it's a little
puzzling as to why Sun is the first to implement the most obvious
scheme. Maybe Intel has been too busy trying to be clever?
I would guess that the performance gain is all relative to an in-order
CPU. As soon as you go to an out-of-order superscalar design the OOO
core finds other ways of hiding memory latency.
Intel has this in-order core it's been fiddling with for lo these many |
years. That in-order core has been the cause of many interesting
papers on latency-hiding techniques. I don't remember seeing this one,
and Intel certainly hasn't implemented it.
| Quote: | Also, Sun is claiming a performance advantage on _commercial
workloads_ (by which I presume they mean transaction processing).
From what I have seen, Intel has cared much more about SPEC-like
workloads and desktop PC workloads in the past --- I suspect that
Sun's hardware may not do as well on such workloads (and Sun may not
care).
Transaction processing is certainly important for Itanium as a proposed |
competitor to Power.
RM |
|
| Back to top |
|
|
Seongbae Park
Guest
|
| Posted:
Thu Jul 14, 2005 4:29 pm Post subject:
Re: A newly published paper on hardware scout thread. |
|
|
Patrick Schaaf <mailer-daemon@bof.de> wrote:
....
| Quote: | Who cares about publishers staying in business, except for the
publishers themselves?
More to the point, what does SUN, as the author of the paper, gain by
supporting this publisher and excluding more casual readership?
best regards
Patrick
|
This is somewhat off-topic on this group so my apology in advance.
But I'd like to give my personal opinion as an answer to this question.
As usual, I don't and can't speak for Sun.
Publishing on the well established journals and conferences
has been the best way to get to the core technical audience,
because most people who have more than "casual" interest in the field
have not only access to those - most quality journals and proceedings
from IEEE or ACM - but regularly check them.
So it is one of the best ways to get their attention,
as well as any future students of the topic
since most of them will start
by scanning through those journals and proceedings.
Although the internet is a relatively easy and low cost channel
to provide any material for free,
the internet made it even more important to get
ideas published in the traditional way -
quality journals and conference proceedings can be
a valuable editorial service for filtering out garbage or subquality works.
However, publishing through traditional paper journals
doesn't preclude providing the paper for free download -
many academics do just that, providing the paper
that's published in the traditional paper journal
as a free download on their own website.
And Sun does similar through docs.sun.com and other means.
If the wider audience wants contents like this particular paper
(and probably if the marketing department thinks it's a good idea),
the technical content of this particular paper may well end up
on somewhere in Sun's websites. This just hasn't happened.
Another beneficial side effect of publishing to a technical journal
is that the papers published on a technical journal tend to have
less content-free marketing-speaks.
--
#pragma ident "Seongbae Park, compiler, http://blogs.sun.com/seongbae/" |
|
| Back to top |
|
|
Maynard Handley
Guest
|
| Posted:
Fri Jul 15, 2005 12:11 am Post subject:
Re: A newly published paper on hardware scout thread. |
|
|
In article <db63tc$9ui$1@news1nwk.SFbay.Sun.COM>,
Seongbae Park <Seongbae.Park@Sun.COM> wrote:
| Quote: | Patrick Schaaf <mailer-daemon@bof.de> wrote:
...
Who cares about publishers staying in business, except for the
publishers themselves?
More to the point, what does SUN, as the author of the paper, gain by
supporting this publisher and excluding more casual readership?
best regards
Patrick
This is somewhat off-topic on this group so my apology in advance.
But I'd like to give my personal opinion as an answer to this question.
As usual, I don't and can't speak for Sun.
Publishing on the well established journals and conferences
has been the best way to get to the core technical audience,
because most people who have more than "casual" interest in the field
have not only access to those - most quality journals and proceedings
from IEEE or ACM - but regularly check them.
So it is one of the best ways to get their attention,
as well as any future students of the topic
since most of them will start
by scanning through those journals and proceedings.
|
Not true. It's a way to make contact with an audience of academics and
academic-like thinkers. It does NOT get you contact with people of a
different mindset; people who may well be very competent engineers,
enthusiastic hobbyists and so on.
You/Sun are welcome to ignore that audience, but IMHO it's pretty stupid
to do so given that it's a lot larger than the academic audience you are
targeting. You/Sun are welcome to secretly pretend that people who don't
follow an academic life-style are stupider and less worthy than you, but
again, IMHO, this is rather blinkered behavior. You/Sun are welcome to
publish 20th century style, but it seems a bit strange that Sun, the
most vocal proponent of company blogs, should be supporting a publishing
model that has been abandonded as inefficient at spreading knowledge by
the physics community (xarchiv), and is in the process of being reworked
to be more public by the biosciences community.
| Quote: | Although the internet is a relatively easy and low cost channel
to provide any material for free,
the internet made it even more important to get
ideas published in the traditional way -
quality journals and conference proceedings can be
a valuable editorial service for filtering out garbage or subquality works.
However, publishing through traditional paper journals
doesn't preclude providing the paper for free download -
many academics do just that, providing the paper
that's published in the traditional paper journal
as a free download on their own website.
And Sun does similar through docs.sun.com and other means.
If the wider audience wants contents like this particular paper
(and probably if the marketing department thinks it's a good idea),
the technical content of this particular paper may well end up
on somewhere in Sun's websites. This just hasn't happened.
|
And by the time it does happen, you will have lost the buzz associated
with the concept. Good for Sun marketing. That's why you're number one
<not>.
| Quote: | Another beneficial side effect of publishing to a technical journal
is that the papers published on a technical journal tend to have
less content-free marketing-speaks.
|
Hmm, I've not noticed that to be a problem with the IBM JRD papers, or
xarchiv papers, or the papers I download from academic's web sites.
Sounds like a rather desperate rationalization.
Yes, in a world where everyone can publish, filters and editing are
needed. Jumping from that to the defense that that means the paper
should not be freely available is a stretch.
|
|
| Back to top |
|
|
|
|
|
|
FAQ
Memberlist
Register
Profile
Log in to check your private messages
Log in