[Q] The memory wall problem: a quantiative approach?
CASTalk.com Forum Index CASTalk.com
Discussion of DSP, FPGA, storage and embedded system.
 
 FAQFAQ   MemberlistMemberlist     RegisterRegister 
 ProfileProfile   Log in to check your private messagesLog in to check your private messages   Log inLog in 
 
Google
 
Web castalk.com
[Q] The memory wall problem: a quantiative approach?

 
Post new topic   Reply to topic    CASTalk.com Forum Index -> Computer Architecture
Author Message
Guest
Posted: Wed Nov 23, 2005 11:47 pm    Post subject: [Q] The memory wall problem: a quantiative approach? Reply with quote
I have been wanting to conduct a quantitative study on the memory wall
problem for recent computer systems. Most papers only provide a
qualitative description of the memory wall problem.

Specifically, I'd like to state something like "processor performance
and memory performance increase by X% and Y%, repsectively, in the past
Y years". There has been studies on the value of X using statistical
method on SPEC benchmarks. I am wondering whether there is any similar
study on the value of Y.

Any pointer or reference is appreciated.

Chung-Hsing
Back to top
Del Cecchi
Guest
Posted: Thu Nov 24, 2005 8:58 am    Post subject: Re: [Q] The memory wall problem: a quantiative approach? Reply with quote
<chunghsu@comcast.net> wrote in message
news:1132768079.089035.94980@g44g2000cwa.googlegroups.com...
Quote:
I have been wanting to conduct a quantitative study on the memory wall
problem for recent computer systems. Most papers only provide a
qualitative description of the memory wall problem.

Specifically, I'd like to state something like "processor performance
and memory performance increase by X% and Y%, repsectively, in the past
Y years". There has been studies on the value of X using statistical
method on SPEC benchmarks. I am wondering whether there is any similar
study on the value of Y.

Any pointer or reference is appreciated.

Chung-Hsing

Depends on what measure you want to use for memory performance. Raw

access/cycle time metrics come from manufacturers web sites and data
sheets. Latency and bandwidth benchmark results should be pretty widely
available on the web and in publications. Detailed information about
DRAM chips is also available from the IEEE in publications such as the
ISSCC digest of technical papers, which go back to the 60's, not that
there were dram chips in the 60's.

del
Back to top
Yogesh
Guest
Posted: Sat Nov 26, 2005 1:15 am    Post subject: Re: The memory wall problem: a quantiative approach? Reply with quote
chunghsu@comcast.net wrote:
Quote:
I have been wanting to conduct a quantitative study on the memory wall
problem for recent computer systems. Most papers only provide a
qualitative description of the memory wall problem.

Specifically, I'd like to state something like "processor performance
and memory performance increase by X% and Y%, repsectively, in the past
Y years". There has been studies on the value of X using statistical
method on SPEC benchmarks. I am wondering whether there is any similar
study on the value of Y.

There is not as much information on the values of Y as much on the
values of X. You will find from previous published research a measure
of the performance gap between X and Y.
A statistical study on the value of X also gives an idea of the
performance of Y. If you are concerned about determining only Y, the
study should involve using benchmarks of your choice and look at the
access pattern of these benchmarks for a memory system, including
studying metrics such as latency, bandwidth and others on the memory
system. This will give you an idea of if/how the memory systems over a
period of time have/have-not provided a performance gain over Y years
if you model all the memory systems during the Y years.

I hope I have understood your questions.!.If this does not help..post
more details.

Yogesh


Quote:
Any pointer or reference is appreciated.

Chung-Hsing
Back to top
Guest
Posted: Thu Dec 01, 2005 1:15 am    Post subject: Re: The memory wall problem: a quantiative approach? Reply with quote
The real goal is to quantify the performance gap between X (processor)
and Y (memory). Yogesh mention that there exist some published research
results on this. Could Yogesh provide more information on that? The
latest such research result I can find out is Dave Patterson's "latency
lags bandwidth" in CACM 2004. Although Dave's article focuses on
something else, the data he provides can be quite useful. Yet, it is
just one data point. Thanks.

Chung-Hsing

Yogesh wrote:
Quote:
chunghsu@comcast.net wrote:
I have been wanting to conduct a quantitative study on the memory wall
problem for recent computer systems. Most papers only provide a
qualitative description of the memory wall problem.

Specifically, I'd like to state something like "processor performance
and memory performance increase by X% and Y%, repsectively, in the past
Y years". There has been studies on the value of X using statistical
method on SPEC benchmarks. I am wondering whether there is any similar
study on the value of Y.

There is not as much information on the values of Y as much on the
values of X. You will find from previous published research a measure
of the performance gap between X and Y.
A statistical study on the value of X also gives an idea of the
performance of Y. If you are concerned about determining only Y, the
study should involve using benchmarks of your choice and look at the
access pattern of these benchmarks for a memory system, including
studying metrics such as latency, bandwidth and others on the memory
system. This will give you an idea of if/how the memory systems over a
period of time have/have-not provided a performance gain over Y years
if you model all the memory systems during the Y years.

I hope I have understood your questions.!.If this does not help..post
more details.

Yogesh


Any pointer or reference is appreciated.

Chung-Hsing
Back to top
Yogesh
Guest
Posted: Sun Dec 04, 2005 9:15 am    Post subject: Re: The memory wall problem: a quantiative approach? Reply with quote
chunghsu@comcast.net wrote:
Quote:
The real goal is to quantify the performance gap between X (processor)
and Y (memory). Yogesh mention that there exist some published research
results on this. Could Yogesh provide more information on that? The
latest such research result I can find out is Dave Patterson's "latency
lags bandwidth" in CACM 2004. Although Dave's article focuses on
something else, the data he provides can be quite useful. Yet, it is
just one data point. Thanks.

Chung-Hsing

I can think of:

"Hitting the Memory Wall: Implications of the Obvious" by Wulf and
Mckee
"The Memory Gap and the Future of High Performance Memories" by Wilkes
"Predicting and Precluding Problems with Memory Latency" by Boland and
Dollas
"Reflections on the Memory Wall" by Mckee (white paper think)

There was a white paper:
"Memory Bandwidth and Performance of Scientific applications" from AMD
which has some numbers on Latency.

I dont think CPU manufacturers show the actual best case memory
latency but you can
get other useful information from the CPU datasheets that might be
helpful in computing
the best case latency with the speed/type of memory system being used.

A survey of memory architectures will also reap useful information.
Also there are lot of white papers that can give lot of information.

Also papers on data prefetching methods can give lot of insight
because they will have a
measure of the latency saved by prefetching and you can relate that
information to the
CPU speeds and help in making a quantitative measurement.

I hope this helps..

Thanks
Yogesh


Quote:
Yogesh wrote:
chunghsu@comcast.net wrote:
I have been wanting to conduct a quantitative study on the memory wall
problem for recent computer systems. Most papers only provide a
qualitative description of the memory wall problem.

Specifically, I'd like to state something like "processor performance
and memory performance increase by X% and Y%, repsectively, in the past
Y years". There has been studies on the value of X using statistical
method on SPEC benchmarks. I am wondering whether there is any similar
study on the value of Y.

There is not as much information on the values of Y as much on the
values of X. You will find from previous published research a measure
of the performance gap between X and Y.
A statistical study on the value of X also gives an idea of the
performance of Y. If you are concerned about determining only Y, the
study should involve using benchmarks of your choice and look at the
access pattern of these benchmarks for a memory system, including
studying metrics such as latency, bandwidth and others on the memory
system. This will give you an idea of if/how the memory systems over a
period of time have/have-not provided a performance gain over Y years
if you model all the memory systems during the Y years.

I hope I have understood your questions.!.If this does not help..post
more details.

Yogesh


Any pointer or reference is appreciated.

Chung-Hsing
Back to top
Dan Koren
Guest
Posted: Sun Dec 04, 2005 9:15 am    Post subject: Re: The memory wall problem: a quantiative approach? Reply with quote
I hope you realize the performance gap is
also problem/algorithm dependent. Without
specifying the problem and the algorithms
one cannot derive meaningful metrics.


dk


<chunghsu@comcast.net> wrote in message
news:1133392384.030127.196420@g14g2000cwa.googlegroups.com...
Quote:
The real goal is to quantify the performance gap between X (processor)
and Y (memory). Yogesh mention that there exist some published research
results on this. Could Yogesh provide more information on that? The
latest such research result I can find out is Dave Patterson's "latency
lags bandwidth" in CACM 2004. Although Dave's article focuses on
something else, the data he provides can be quite useful. Yet, it is
just one data point. Thanks.

Chung-Hsing

Yogesh wrote:
chunghsu@comcast.net wrote:
I have been wanting to conduct a quantitative study on the memory wall
problem for recent computer systems. Most papers only provide a
qualitative description of the memory wall problem.

Specifically, I'd like to state something like "processor performance
and memory performance increase by X% and Y%, repsectively, in the past
Y years". There has been studies on the value of X using statistical
method on SPEC benchmarks. I am wondering whether there is any similar
study on the value of Y.

There is not as much information on the values of Y as much on the
values of X. You will find from previous published research a measure
of the performance gap between X and Y.
A statistical study on the value of X also gives an idea of the
performance of Y. If you are concerned about determining only Y, the
study should involve using benchmarks of your choice and look at the
access pattern of these benchmarks for a memory system, including
studying metrics such as latency, bandwidth and others on the memory
system. This will give you an idea of if/how the memory systems over a
period of time have/have-not provided a performance gain over Y years
if you model all the memory systems during the Y years.

I hope I have understood your questions.!.If this does not help..post
more details.

Yogesh


Any pointer or reference is appreciated.

Chung-Hsing
Back to top
Alexi Tekhasski
Guest
Posted: Wed Dec 14, 2005 9:15 am    Post subject: Re: The memory wall problem: a quantiative approach? Reply with quote
"Yogesh" <yogesh.kinikar@gmail.com> wrote in message
news:1133674011.055938.13490@g49g2000cwa.googlegroups.com...
Quote:

chunghsu@comcast.net wrote:
The real goal is to quantify the performance gap between X (processor)
and Y (memory). Yogesh mention that there exist some published research
results on this. Could Yogesh provide more information on that? The
latest such research result I can find out is Dave Patterson's "latency
lags bandwidth" in CACM 2004. Although Dave's article focuses on
something else, the data he provides can be quite useful. Yet, it is
just one data point. Thanks.

Chung-Hsing

I can think of:

"Hitting the Memory Wall: Implications of the Obvious" by Wulf and
Mckee
"The Memory Gap and the Future of High Performance Memories" by Wilkes
"Predicting and Precluding Problems with Memory Latency" by Boland and
Dollas
"Reflections on the Memory Wall" by Mckee (white paper think)

There was a white paper:
"Memory Bandwidth and Performance of Scientific applications" from AMD
which has some numbers on Latency.



I would also go ahead and suggest my own humble (and mostly
unappreciated) contribution to the problem of quantification
of the memory wall issue,
http://home.austin.rr.com/bah/
with some data from different workloads.

Cheers,
- aap

(P.S. Unfortunately, the page suffered some screwup during my recent
attempt to clean it up, so some pictures do not show up; sorry
about that).
Back to top
 
Post new topic   Reply to topic    CASTalk.com Forum Index -> Computer Architecture All times are GMT
Page 1 of 1

 
 You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum




VoIP Electronics Powered by phpBB