'TODO'에 해당되는 글 26건

PROGRESS REPORT (8/10)

Lab.work 2009. 8. 11. 01:02
■ THIS WEEK  (8/4~8/10)

□ Phoenix 2.0
I ran pheonix on gems (simics + ruby).
I just tested wordcount and pca.
Both applications still have a few assertion error to fix, but they worked.
But, with ruby, it takes too much time to watch the process.

□ Booksim 2.0
i changed the Bernoulli function and made bimodal injection function.
the result is as below.
(booksim_uniform_bimodal.pdf)

And I draw the traffic patterns correctly.
But we still need to lookup booksim code and analyze each traffic pattern.

□ sigMP
A Novel Cache Architecture with Enhanced Performance and Security (MICRO 2008)


■ NEXT WEEK  (8/11~8/17)

□ Phoenix 
run all applications on gems with different configurations.

□ Booksim 2.0
read code in detail and understand the process. 
test and compare the result, and analysis the characteristic of each traffic pattern.


'Lab.work' 카테고리의 다른 글

PROGRESS REPORT (8/26)  (0) 2009.08.26
PROGRESS REPORT (8/17)  (0) 2009.08.17
PROGRESS REPORT (8/3)  (0) 2009.08.04
PROGRESS REPORT  (0) 2009.07.28
PROGRESS REPORT  (0) 2009.07.21
블로그 이미지

민둥

,

PROGRESS REPORT (8/3)

Lab.work 2009. 8. 4. 10:27
■ THIS WEEK  (7/28~8/3)

□ Phoenix 2.0
I analyzed the result of histogram.
average task processing time are increases as # of cores increase.
but, it does not increase dramatically as compared with wordcount.
(new_result_histogram.pdf)

□ Booksim 2.0
I read the book, and drew each traffic patterns.
I'm not sure this is correct, but there are some strange things.
Bit complement (D_i = ~S_i) and Bit reverse (D_i = S_b-i-1) look exactly the same.
for Shuffle(D_i = S_i-1 mod b) and Bit rotation(D_i = S_i+1 mod b), only the traffic direction is different.
and when mesh size is 4x4, transpose and neighbor are the same.
(traffic_patterns.jpg)

we are checking these features with the results.

□ sigMp
no sigMp seminar this week.


■ NEXT WEEK  (8/4~8/10)

□ Phoenix 
run a couple of applications more and compare the result with wordcount.

□ Booksim 2.0
test and compare the result, and analysis the characteristic of each traffic pattern.

□ sigMp
Jaehong asked me to cover for his turn next week (8/14).
so please introduce me a paper for the presentation.

'Lab.work' 카테고리의 다른 글

PROGRESS REPORT (8/17)  (0) 2009.08.17
PROGRESS REPORT (8/10)  (0) 2009.08.11
PROGRESS REPORT  (0) 2009.07.28
PROGRESS REPORT  (0) 2009.07.21
PROGRESS REPORT  (0) 2009.07.07
블로그 이미지

민둥

,

PROGRESS REPORT

Lab.work 2009. 7. 28. 11:54
■ THIS WEEK  (7/21~7/27)

□ Phoenix 2.0

I found that thier scheduler does not work actually.
so they tested using threads which were not fairly distributed to cores.

the threads were mapped core_id = thread_id
therefore, when # of threads <=8, they used only one cpu.
when # of threads=16, they used 2 cpu, when # of threads=32, # of cpu is 4 and so on.

I fixed the code to distribute the threads to cores fairly.
the result is very different and the performance is also very higher than before.
(new results)

□ Booksim 2.0

we had a discussion about the traffic patterns and the expected results.
simple traffic pattern like neighbor is relatively easy to predict but others are not so easy.
we think that we have a misunderstanding of the uniform traffic and need more reseach.
(files of 4 traffic patterns)


■ NEXT WEEK  (7/28~8/3)

□ Phoenix 
re-plot the graph of some other applicatoins

□ Simics & Gems
I think now i should try to run phoenix on Gems.

□ Booksim 2.0
discuss more about the traffic pattern and expeted result.
test more with different config files(using different delay).

'Lab.work' 카테고리의 다른 글

PROGRESS REPORT (8/10)  (0) 2009.08.11
PROGRESS REPORT (8/3)  (0) 2009.08.04
PROGRESS REPORT  (0) 2009.07.21
PROGRESS REPORT  (0) 2009.07.07
6월 18일 발표논문  (0) 2009.07.01
블로그 이미지

민둥

,

PROGRESS REPORT

Lab.work 2009. 7. 21. 10:39
■ THIS WEEK  (7/15~7/20)

□ Phoenix 2.0

about degradation after 32 threads:  
first, I conjectured that the reason was skewed distribution of words.
so I randomly generated input file using a-z and white space.
but there still exists slight degradation.

now, I divided the start_worker function into 7 parts and find the degradation part.
here I attached my work. 

□ Booksim 2.0

the simulation script is still running.
the result file using uniform traffic is attached.
I will analysis the result when other data come out.


■ NEXT WEEK  (7/21~7/27)

□ Phoenix 
Test more with huge data set and re-plot the speedup graph.
Find way to determine working set size and L2.

□ Simics & Gems
a little later..

□ Booksim 2.0
Plot all graph using other traffics and analysis the result.

'Lab.work' 카테고리의 다른 글

PROGRESS REPORT (8/3)  (0) 2009.08.04
PROGRESS REPORT  (0) 2009.07.28
PROGRESS REPORT  (0) 2009.07.07
6월 18일 발표논문  (0) 2009.07.01
6월 29일 월요일  (0) 2009.06.29
블로그 이미지

민둥

,

PROGRESS REPORT

Lab.work 2009. 7. 7. 11:38
■ THIS WEEK  (6/29~7/6)

□ Phoenix 2.0
I re-ploted the result graph with the x-axis.
There is no degredation after 32 nodes when data set is huge.
Optimal chunk size is dependent on the application.
□ Booksim 2.0
I tested 2D mesh network with different number of virtual channel and buffer size.

■ NEXT WEEK  (7/7~7/13)

□ Phoenix 
Test more with huge data set and re-plot the speedup graph.
Find way to determine working set size and L2.

□ Simics & Gems
run and test PARSEC on Simics and Gems.

□ SigMP
SigMP seminar will be on Thursday. 

'Lab.work' 카테고리의 다른 글

PROGRESS REPORT  (0) 2009.07.28
PROGRESS REPORT  (0) 2009.07.21
6월 18일 발표논문  (0) 2009.07.01
6월 29일 월요일  (0) 2009.06.29
PROGRESS REPORT (6/29)  (0) 2009.06.29
블로그 이미지

민둥

,

6월 29일 월요일

Lab.work 2009. 6. 29. 22:21
■ Sun Machine 

케이블 가격 알아보기
혹시 자주 다운된다면 다시 OS 깔아야하나.

■ Phoenix

check ** L2가 working set size를 포함하기 못할때..!!

Phoenix on Simics!ㅠ_ㅠ
일단 실제 기계에서 다 돌려보고 Simics에서 해보삼

wordcount가 왜 초반에는 super linear speedup 인데 다음에는 떨어지는걸까
wordcount working set이랑 L2 size 확인
다른 application에서도 chuck size가 어떤 영향을 미치는지

Speedup 그래프 다시 그리기
chunk size가 정확히 어떤 영향을 미치는지 + best chunk size적용

linux x86랑 sun sparc이랑 무슨 차이? 왜??
hyper threading? super 어쩌고 threading? throughput이 더 중요한지 연산을 하는데 최적화 되어있는지 등등
+ thread끼리 공유하는 cache등등의 문제
sun 같은 경우는 모두 on chip communication이라서!?
linux같은 경우는 하나의 board에서... 그래서?

sun system config 찾아보기

■ Booksim

목요일날 다시 얘기!
Booksim에 대해서 다시 설명해주심
topology에 대해서 간단히 찾아보고, code도 대충 보기!

networks.cpp
iq_router.cpp
가장 highlevel은 trafficmanager.cpp <= 여기서 모든게 실행

"wait_for_tail_credit" config 살펴보기
booksim_config.cpp

■ Gems

목요일전에 gems돌려보기

-------------------------------------

목요일 한국시간 8시 다시 미팅


'Lab.work' 카테고리의 다른 글

PROGRESS REPORT  (0) 2009.07.07
6월 18일 발표논문  (0) 2009.07.01
PROGRESS REPORT (6/29)  (0) 2009.06.29
PROGRESS REPORT (6/23)  (0) 2009.06.23
PROGRESS REPORT (6/15)  (0) 2009.06.16
블로그 이미지

민둥

,

PROGRESS REPORT (6/29)

Lab.work 2009. 6. 29. 16:19
■ THIS WEEK  (6/24~6/28)

□ Sun SPARC machine
Sun SPARC machine does not work since friday evening.
OS was downed, but we could't find the reason.
the directory where the OS setup is too small, I think it might be the reason.
anyway, It works fine now.

□ Phoenix 
the results fils is updated. 
speedups are recalculated based on the result when num_proc is 1.
Phoenix_version2.0.pdf

□ Booksim 2.0
I downloaded the Booksim simulator and test with example files.
□ Simics & Gems
김창대 student setup PARSEC on Simics except raytrace, vips applications and cmake tool.
I am following with reference to the wiki page.



■ NEXT WEEK  (6/29~7/6)

□ Phoenix 
Find the reason why wordcount speedup decreases when num_proc = 64

□ Simics & Gems
run and test PARSEC on Simics and Gems.

'Lab.work' 카테고리의 다른 글

6월 18일 발표논문  (0) 2009.07.01
6월 29일 월요일  (0) 2009.06.29
PROGRESS REPORT (6/23)  (0) 2009.06.23
PROGRESS REPORT (6/15)  (0) 2009.06.16
6월 11일 발표논문  (0) 2009.06.15
블로그 이미지

민둥

,

PROGRESS REPORT (6/23)

Lab.work 2009. 6. 23. 22:03
■ THIS WEEK  (6/16~6/23)

□ Phoenix 
The results fils is attached. 
Phoenix_version2.0.pdf

□ SigMP seminar (6/18)
An OS-Based Alternative to Full Hardware Coherence on Tiled CMPs (HPCA 2008) (by 김대훈)
PARSEC vs. SPLASH-2: A Quantitative Comparison of Two Multithreaded Benchmark Suites (by 신민정) 

□ Simics & Gems
compiled PARSEC binary file on linux machine.



■ NEXT WEEK  (6/23~6/29)

□ SigMP seminar
no sigMP seminar

□ Simics & Gems
run and test PARSEC on Simics and Gems.


'Lab.work' 카테고리의 다른 글

6월 29일 월요일  (0) 2009.06.29
PROGRESS REPORT (6/29)  (0) 2009.06.29
PROGRESS REPORT (6/15)  (0) 2009.06.16
6월 11일 발표논문  (0) 2009.06.15
6월 4일 발표 논문  (0) 2009.06.09
블로그 이미지

민둥

,

PROGRESS REPORT (6/15)

Lab.work 2009. 6. 16. 10:47
■ THIS WEEK  (6/9~6/15)

□ Phoenix 

(1) Unit size
I checked the unit_size of each application.
They tested unit_size with 4KB, 16KB, 64KB, and 128KB in the paper, but the default size of each application is as below.

they calculated chunk size using the unit size and L1 cache size. (chunk_size = unit_size / L1_cache_size)
but their code cannot get L1 cache size from system environment, and always use default size as 64KB.
I will get the real cache size and make a graph again. 

(2) # of threads
I am not sure whether the threads are divided equally among processors or not. 
I tried to find out the way, but couldn't find the system command on Solaris.

(3) Sequential time
Sequential time results of each application are almost same, but they are not perfectly equal.
i think it's because the machine state is a little different each time.

(4) One processor speedup 
the result graph is attached at the file. 
the processing times of mapreduce and pthreads are similar to the sequential, 
but mapreduce performed better for some implementation (ex. wordcount, reverseindex). 

□ SigMP seminar (6/11)
Coordinated Management of Multiple Interacting Resources in Chip Multiprocessors: A Machine Learning Approach (Micro 2008) 
Bandwidth adaptive snooping (HPCA 2002)

□ Simics & Gems
Study and get used to Simics. run SPLASH2 and PARSEC.



■ NEXT WEEK  (6/16~6/22)

□ SigMP seminar (6/18)
An OS-Based Alternative to Full Hardware Coherence on Tiled CMPs (HPCA 2008) (by 김대훈)
PARSEC vs. SPLASH-2: A Quantitative Comparison of Two Multithreaded Benchmark Suites (by 신민정) prepare Seminar

□ Simics & Gems
Setup Gems simulator. run SPLASH2 and PARSEC.

'Lab.work' 카테고리의 다른 글

PROGRESS REPORT (6/29)  (0) 2009.06.29
PROGRESS REPORT (6/23)  (0) 2009.06.23
6월 11일 발표논문  (0) 2009.06.15
6월 4일 발표 논문  (0) 2009.06.09
PROGRESS REPORT (6/8)  (0) 2009.06.09
블로그 이미지

민둥

,

PROGRESS REPORT (6/8)

Lab.work 2009. 6. 9. 00:09
■ THIS WEEK - WHAT I DID (6/2~6/8)

SigMP seminar (6/4)
Self-Optimizing Memory Controllers: A Reinforcement Learning Approach (ISCA 2008)
Virtual Hierarchies to Support Server Consolidation (ISCA 2007)

Phoenix
I wanted to test performance using 4 cores (vc105 machine), but there was a problem.
To change a cpu option, the machine has to be rebooted with maxcpus boot options.
I asked the manager of the server to reboot with the options, because I couldn't do that remotely.
The manager said, he is still finding a way to do that.

instead of waiting,
I tested Phoenix version 1.0 on Sun T2 machine (cst5210 machine) with 8 cores (64 threads)
the result graph is in the file attached.
for every applications, the results using mapreduce are no better than p-threads.
compare to CMP speedup in the paper, using 8 cores (32 threads), 
even reverse_index and word_count didn't perform well using mapreduce.

I think when we have many workers, the overhead of mapreduce increases.
the approach in the paper was fresh, but I'm not sure it would be good for the future multicore/multiprocessor systems.

Simics & Gems
since the nodes from Prof. Moon do not have enough space, I got an account of pcal6 machine from a student of Prof. Huh, 
now I am sharing pcal6 machine with JeongSeob.
I downloaded and installed Simics 3.0 on pcal6 again.

SigSimics meeting was on Thursday.
I applied for studying PARSEC benchmark together.
I am trying to figure out the characteristic, and understand the difference between PARSEC and SPLASH2.

Paper
The SPLASH-2 Programs: Characterizationand Methodological Considerations
The PARSEC Benchmark Suite: Characterization and Architectural Implications
PARSEC vs. SPLASH-2: A Quantitative Comparison of Two Multithreaded Benchmark Suites on Chip-Multiprocessors



■ NEXT WEEK - WHAT TO DO (6/9~6/15)

SigMP seminar (6/11)
Coordinated Management of Multiple Interacting Resources in Chip Multiprocessors: A Machine Learning Approach (Micro 2008) 
Bandwidth adaptive snooping (HPCA 2002)

Simics & Gems
Study and get used to Simics. Setup Gems simulator.
run SPLASH2 and PARSEC.

'Lab.work' 카테고리의 다른 글

6월 11일 발표논문  (0) 2009.06.15
6월 4일 발표 논문  (0) 2009.06.09
PROGRESS REPORT (6/1)  (0) 2009.05.31
5월 27일 수요일  (0) 2009.05.27
PROGRESS REPORT (5/22)  (0) 2009.05.22
블로그 이미지

민둥

,

공지사항

최근에 받은 트랙백

글 보관함

달력

«   2025/05   »
일 월 화 수 목 금 토
1 2 3
4 5 6 7 8 9 10
11 12 13 14 15 16 17
18 19 20 21 22 23 24
25 26 27 28 29 30 31

티스토리툴바