|
D**0
发帖数: 2048 | 2 ☆─────────────────────────────────────☆
Muchistrue (HelloThere) 于 (Tue Oct 29 02:35:26 2013, 美东) 提到:
赞👍！
☆─────────────────────────────────────☆
rainhorse (rainhorse) 于 (Tue Oct 29 02:42:00 2013, 美东) 提到:
Support!!!
☆─────────────────────────────────────☆
cycle (wolfpack) 于 (Tue Oct 29 02:44:58 2013, 美东) 提到:
我要拿横幅，没法举牌子，一个中年白人女士主动要求帮我举牌子，后来她要走
的时候还一再和我说你们干得好。路上也有行人问起我们抗议的原因，听我们
解释后都表示非常支持。我们的抗议是有广泛的群众基础的，我们应该让更多
的人了解和加入。
☆─────────────────────────────────────☆
redscarf (... 阅读全帖 |
|
|
|
s******s
发帖数: 190 | 5 I wonder how much speedup you can get after switching to the high-end shoes.
Does it really make that much difference? |
|
w*******e
发帖数: 2651 | 6 Wow! 35mi/wk is very good mileage. Recovery is important too.
A lot of ice and warm cycles probably would help
speedup the recovery. Take care! |
|
w*******e
发帖数: 2651 | 7 Don't know exactly how you would call it, but I would
use a lot of ice (or a bag of frozen peas) to ice it.
It will help speedup recovery.
BTW, are you a heel striker? |
|
w*******e
发帖数: 2651 | 8 That's one of the reason for runners to drink some coffee
before a race start. It would help speedup heartbeat to
prime your cardio for the running pace. |
|
c*********n
发帖数: 2120 | 9 Maybe you are right. I did try to speedup after mile 10 and I am just
learning 2-2 breath. But I thought these are right ways...
Did you find some solutions? |
|
c*********n
发帖数: 2120 | 10 To run CIM turns out to be such a blessing thing for me. Xiaolong and
yuanyuan are just so sweet to make my 3 days in Sacramento the most comfort,
and they even covered all my food-:)It was very happy for me to meet with
so many other runners for whom now I can connect their IDs with their smile-
close! I almost asked myself why I have waited such a long time to meet
yours-:)
During the race, I realized I did not have much confidence for BQ, maybe
only 50%-:) It was precious to have baiji pace m... 阅读全帖 |
|
s*********r
发帖数: 3934 | 11
对了，谢谢你，不过还是一堆noise，升级了也没用，我后来买了GB－10tascam的还挺
好用的，唯一的缺点就是slowdown和speedup都是10％，太他爹的多了，3％还可以。 |
|
y*b
发帖数: 3190 | 12 http://www.43rumors.com/adobe-announces-the-new-lightroom-cc/
"Adobe announced the new Lightroom CC (details at Adobe here) and Lightroom
6 (details ad Adobe here). It adds plenty of features like Facial
Recognition, new panorama, new brushes for graduated and radial filters,
more slideshow options, HDR (high dynamic range), as well as significant
speedups (up to 1000% faster!). Also the mobile APP of Lightroom got
improved with segmented view, easier rating method, improved crop and copy
and pa... 阅读全帖 |
|
S*********r
发帖数: 5693 | 13 哈哈哈哈
400 www.fuckqq.com have exceed the daily limited
To speedup www.fuckqq.com visiting please use sub website directly !
not using URL forward.
How to setup and IIS or Apache sub website in VDNS
This page will redirect to www.360.cn 10 seconds later
Go to http://www.360.cn NOW |
|
g****t
发帖数: 39 | 14 Thank you. I noticed these conferences also.
Any good one on analytical models? Typical papers on Grid are as follows:
1. a framework, preliminary implementation, some application experiments, show
several tables and figures, then give some intuitive thinking
2. an algo, with complex and fancy flowchart, some experiments, a little bit
on why and how
Any papers show the upper bound on the performance limit, speedup/efficiency,
analytical models, stochastic analysis etc. I noticed one stochastic p |
|
g****t
发帖数: 39 | 15 Grid is not only for fun :). It can be though, such as mass game platform with
millions of players.
Don't pay too much attention to rumors. He did great contribution already. We
have to ask ourselves what we can do?
show
bit
speedup/efficiency,
paper. I |
|
j******g
发帖数: 436 | 16 现在高端的CPU大概是100gflops/S，内存是几十GB/S，GPU是1-2Tflops/S，150GB/S。
所以针对那些可以优化的algorithm，GPU的speedup算10x比较靠谱。至于优化的难度就
像你说得cache line，SSE，在CPU上的优化一样很难。GPU的领域是有局限性。主要还
是一些算法。不过现在反而是GPU越来越多CPU的技术，因为GPU比较简单，可以复杂化
。现在GPU变化特别大。反而是CPU已经很复杂了。
如C |
|
j******g
发帖数: 436 | 17 现在高端的CPU大概是100gflops/S，内存是几十GB/S，GPU是1-2Tflops/S，150GB/S。
所以针对那些可以优化的algorithm，GPU的speedup算10x比较靠谱。至于优化的难度就
像你说得cache line，SSE，在CPU上的优化一样很难。GPU的领域是有局限性。主要还
是一些算法。不过现在反而是GPU越来越多CPU的技术，因为GPU比较简单，可以复杂化
。现在GPU变化特别大。反而是CPU已经很复杂了。
如C |
|
s****d
发帖数: 56 | 18 Data mining is born from the database area and focuses how to speedup a
simple algorithm (e.g., Apriori, FP-tree, DBScan), while machine learning is
born from the AI area and focuses on how to build a more accurate model
from the known data (e.g., neural network, Bayes, SVM). |
|
c******a
发帖数: 600 | 19 我现在做的CUDA程序在三台机器上跑，8600M(自己的laptop,在家编程调试方便），
9800GX2和Tesla（两个板）。就性能而言，9800GX2已经有挺理想的speedup了，Tesla
还可以进一步提高2-3倍（当然，取决于应用）。不过总体来说性价比未必划算 |
|
b********p
发帖数: 875 | 20 ms atom 330 speedup很高啊, 高过2
ION显卡能里比集成的高多了
用cinebench测的, 数据来自于
http://benchmarkreviews.com/index.php?
option=com_content&task=view&id=334&Itemid=69&limit=1&limitstart=9
ION + N330: 1184 (openGL) / 538(single) / 1534(Duo)
和其他比较
Acer1410/SU2300/4500 585 / 1376 / 2574
T2600 / 945 584 / 1919 / 3613
T2600/ Go7400 2185 / 1925 / 3628 |
|
kn
发帖数: 2446 | 21 core的数目就是看Stream processor的数目。
至于GPU和CPU的比较，Intel的人发了篇paper，叫做“Debunking the 100X GPU vs. C
PU Myth”。今年6月ISCA的新paper。
如果CPU上的程序用SSE指令优化和并行编程，实际speedup没有传说中那么巨大。一般说
GPU比CPU加速几十上百倍，大多是GPU厂商的市场宣传。用GPU的并行程序、特殊指令和
CPU的普通c程序相比，有失公平。
倍？ |
|
y**b
发帖数: 10166 | 22 我这是从dell买的，狠算了四年了，没啥问题。
一般认为hyperthread对并行浮点计算没啥明显帮助，
而且影响speedup/efficiency评价效果，
都关掉了。 |
|
y**b
发帖数: 10166 | 23 我这是从dell买的，狠算了四年了，没啥问题。
一般认为hyperthread对并行浮点计算没啥明显帮助，
而且影响speedup/efficiency评价效果，
都关掉了。 |
|
y**b
发帖数: 10166 | 24 内存这个一般倒是都会做些分析，毕竟并行计算并不只是speedup，内存够用也非常重
要。有些计算对内存需求极大，非得到多个节点上均分一下才能算得动，典型的如三维
边界元方法。
好的scalability很重要的一点就是problem size增大的时候节点内存无需增加或者只
需缓慢增加，而且还能保持iso-efficiency。对不同的具体问题这种scalability都是
不同的，一般需要推演相关公式，如三维差分流动和三维离散元就有不同的
scalability，做深度耦合并行的时候就得很注意。
但是cache这东西实在很敏感，好像也没见到啥理论分析，不过我是搞应用的这方面知
之不深。 |
|
E***e
发帖数: 3430 | 25 MATLAB GPU is a joke.
If you want the speedup but don't want to learn CUDA, try Arrayfire. |
|
g****g
发帖数: 1828 | 26 从远程机器往自己笔记本拷东西。最后有个小error。但我查看了一下，貌似大多数文
件都考成功了。error是因为什么呢。
sent 28356 bytes received 10150005 bytes 452371.60 bytes/sec
total size is 36899505 speedup is 3.63
rsync error: some files/attrs were not transferred (see previous errors) (
code 23) at main.c(1526) [generator=3.0. |
|
g****g
发帖数: 1828 | 27 嗯，这次好像没问题。 文件没有做任何改变。
receiving file list ...
1473 files to consider
deleting cpvnote/figs/plots/PreTag (Case Conflict 1)/CVS/
deleting cpvnote/figs/plots/PreTag (Case Conflict 1)/
cpvnote/figs/plots/
cpvnote/figs/plots/Pretag/
cpvnote/figs/plots/Pretag/CVS/
cpvnote/figs/plots/Pretag/CVS/Entries
2 100% 1.95kB/s 0:00:00 (xfer#1, to-check=170/1473)
cpvnote/figs/plots/Pretag/CVS/Repository
55 100% 53.71kB/s 0:00:00 (xfer#2, to-check=169/1473)
cpvnote/figs/plots/Preta... 阅读全帖 |
|
y****i
发帖数: 156 | 28 download psyco and usually I do psyco.full() to speedup |
|
Z****e
发帖数: 2999 | 29 数据量不够大？MPI有overhead，如果数据量小的话额外的overhead赶不上speedup就会
变得更慢 |
|
I******c
发帖数: 163 | 30 GPU上跑程序得到的speedup有时候取决于gpu内存的带宽。现在gpu内存的带宽虽然比
cpu的快，但是有没有快50倍？ 另外，我个人理解gpu的带宽是指在最佳情况下，比如
说所有内存的读写都是coalesce了，这个在实际中并不是都可以实现。 |
|
v***r
发帖数: 1046 | 31 needs parallel programming, and it is not trivial work, the performance
depends on your algorithm, and the speedup is usually less than linear. |
|
j******g
发帖数: 436 | 32 我的看法是指针操作是很不好的行为。不过也不重要，很多时候能够省一半的开发时间
，比2x speedup重要的多了。就比如你举得例子吧，你也不会真正evaluate对整个系统
有多大的影响，去看instructions究竟发生了什么。 |
|
w***g
发帖数: 5958 | 33 一个做纯计算的C++程序，用openmp并行化，发现用root跑比正常快25% ，怎么回事？
UPDATE:真实原因是我设了OMP_NUM_THREADS=6的环境变量，sudo的时候这个环境变量没
有带进去。CPU是6核12个线程，那25%的speedup是hyperthreading造成的。 |
|
w***g
发帖数: 5958 | 34 牛！虽然发现不是这个原因造成的，还是惊异于魏老师能这么快地给出一个可能的原因。
真实原因是我设了OMP_NUM_THREADS=6的环境变量，sudo的时候这个环境变量没有带进
去。CPU是6核12个线程，那25%的speedup是hyperthreading造成的。 |
|
S*A
发帖数: 7142 | 35 手写其实更容易维护。看看 gcc 转换的时候是如何说的，
YACC 把代码切成一小段一小段的，中间还有个
黑盒子的状态树，其实维护更加困难。C 其实不是
严格 LLK 语法的。
BTW 注意到手写后速度还轻微提高了没有？
http://gcc.gnu.org/wiki/New_C_Parser
Benefits:
Although timings showed a 1.5% speedup, the main benefits are
facilitating of future enhancements including: OpenMP pragma support; lexing
up front for C so reducing the number of different code paths; diagnostic
location improvements (and potentially other diagnostic improvements);
merging cc1/cc1obj into a single executable with runti... 阅读全帖 |
|
w***g
发帖数: 5958 | 36 如果S是固定的不需要支持插入删除，并且S的取值范围确实很大，可以用perfect hash
，用大量离线计算时间换一点speedup。
你这个总得有一些约束才好优化。如果就一个generic的问题，std::set或者std::
unorderd_set基本上就是最好的了。
S |
|
h*******n
发帖数: 82 | 37 你这种人技术上一说不过别人就拿钱来说事，我说JAVA不适合做游戏，你就说
MINECRAFE卖了2BILLION，你怎么不说同样时代的语言，人家C++的游戏赚了多少，一个
call of duty就超过11个BILLION.
http://www.playstationlifestyle.net/2015/02/06/activision-call-
这么多年来从第一个3D游戏DOOM，到雷神，星际争霸，反恐精英，到魔兽争霸，魔兽世
界，GTA .......哪一个伟大的游戏不是用C++做的，反观这么多年来，除了MINECRAFT
，你再举出一个JAVA做的成功的游戏来。而且我用什么语言都能做MINECRAFT出来，但
是我担保你用JAVA连个10年前的魔兽世界都做不出来。 游戏领域如果C++是100000，
JAVA连0.00000001都不到。游戏最重要靠什么-显卡，你去看看nvidia，ATI之类的招不
招JAVA的人(网站管理员除外).
OK,回到asm.js, 你老是说javascript dynamic type怎么可能比的上JAVA static type
，你研究过as... 阅读全帖 |
|
s******u
发帖数: 501 | 38 跟单线程运行比较，280个线程也就只有50-60x的speedup |
|
d******c
发帖数: 2407 | 39 The Landscape of Parallel Computing Research - UC Berkeley EECS
1. Old CW: Power is free, but transistors are expensive.
· New CW is the “Power wall”: Power is expensive, but transistors are “
free”. That
is, we can put more transistors on a chip than we have the power to turn on.
2. Old CW: If you worry about power, the only concern is dynamic power.
· New CW: For desktops and servers, static power due to leakage can be 40%
of
total power. (See Section 4.1.)
3. Old CW: Monolithic uniprocessors ... 阅读全帖 |
|
x******g
发帖数: 3952 | 40 firefox memory problem never goes away completely.
it is better nowadays than 2004.
but you'll still notice it speedup a lot once you restart firefox.
all chornical memory footprint growth might not be called memory leak.
but they are equally bad whatever you call it.
problem, |
|
p**********t
发帖数: 2636 | 41 啥叫edit distance啊
pair-wise alignment么，local还是global
还有你要bench mark啥，local/global alignmeng各自都有杂七杂八各种speedup |
|
f***n
发帖数: 6 | 42 Just for fun, I run bench.m on my Barton
System:
Barton @ 1.4GHz (reason: my memory is PC2100, so Barton downspeed :-( )
DDR ram 512MB
GF 4 MX440.
Matlab 6.1
Win2k pro.
LU decomp: 0.51s
Then I replace the ATLAS library with the one claimed to be optimized for
AthlonXP
(see here: http://www.econ.ucsd.edu/~ksheppar/matlab/MatlabAtlas.htm)
However, there is NO speedup observed, LU costs even 0.71s!!!
So I switch back to default ATLAS)
I am also curious that if any existent package can use SSE2's a |
|
j**u
发帖数: 6059 | 43 藕的博士论文程序只用了4个左右的很简短的subroutine，就完成了“并行计算”，效
果是：8个processors，速度是8倍；
这个是你写的还是转贴的?很有意思.开始我还以为你自己开发了并行计算的东西，结果
还是用的mpi。其实那些程序用fortran一样很容易实现，不知道贴这个对你想说明的观
点有什么作用?另外你的speedup最高，是因为processor之间的commonication很少，这
个和语言没有任何关系，是和你的问题直接相关的。
10 |
|
