x*********n
发帖数: 28013 | 1 这个配置有什么问题吗？
interface GigabitEthernet0/0
description **** Fiber interface ****
bandwidth 150000
no ip address
ip pim sparse-dense-mode
load-interval 30
duplex auto
speed auto
media-type sfp
no negotiation auto
service-policy output QOS
!
interface GigabitEthernet0/0.70
description **** Fiber interface ****
bandwidth 150000
encapsulation dot1Q 70
ip address wan 255.255.255.248
no ip redirects
no ip unreachables
no ip proxy-arp
谢谢！
! |
|
x*********n
发帖数: 28013 | 2 在FA下面有duplex和speed，
现在Gi下面speed nonegoation了，duplex对应哪个command呢？
客户从FA转到Gi，速度反而从100Mbps变成了50。
会是什么原因呢？
description ***MPLS DATA NETWORK***
ip address 10.3.30.1 255.255.255.128
ip accounting output-packets
ip pim sparse-dense-mode
ip policy route-map To-Exchange
load-interval 30
no negotiation auto |
|
d****n
发帖数: 12461 | 3 多核看来只对设计好的函数有用，例如LU。奇怪的是开了多核以后fft和sparse速度反
而变慢了。
最终还是拼的主频，alu和fpu的性能。 |
|
l***h
发帖数: 392 | 4 本来不想回的，GPU能不能用，看个人。
但不是听起来就好听吧。我自己就在ubuntu下写matrix 相关的程序。
对角化，和sparse matrix multi的东西很成熟了。
我自己家得电脑测试 9800GT比我的Q9400快不是一点两点。价格是 75 vs 170。
至于MD的东西，GPU就更比CPU好得多了，相关得paper你去google一下就有好多。
（GPU，Folding@HOME，openMM）
这个双精度提高了8倍，还是很不错得。
当然不是这个现在就多么牛B了，只是说这个是个趋势，门槛会越来越低。到是
compiler
先进了，就可以按x86一样直接写东西了，不是很好么。
技术总是要进步的。这一步NV对不对不好说，不过对HPC来说觉得是limestone。
等技术成熟了，用GPU powerhouse 替代 纯cpu得cluster是必然得。
|
|
a*s
发帖数: 425 | 5 因为本来这个就hyper的不行
gpu不是通用cpu，
具体看应用
具体到应用，又要看implemention
我觉的gpu有绝对优势的，就是图像处理，还有那些只需要简单运算，没有很多逻辑分
支的计算
那种matrix operation
对于sparse matrix，cpu不会差到哪里，对于dense matrix，gpu自己也会有问题 |
|
|
l*******1
发帖数: 866 | 7
3.
这个要么是主板，要么是power management问题。前面不有人说Ryzen by default是
balanced mode 而不是 performance mode 么？disk IO 这种 sparse short bursts
workload应该受的影响蛮大的。 |
|
l*******1
发帖数: 866 | 8
3.
这个要么是主板，要么是power management问题。前面不有人说Ryzen by default是
balanced mode 而不是 performance mode 么？disk IO 这种 sparse short bursts
workload应该受的影响蛮大的。 |
|
z*********e
发帖数: 10149 | 9 100K * 100K * 8Byte(double) = 80GB?
这么大的矩阵难道不是一般都是sparse的，你这么强行存不合理吧 |
|
c*****t
发帖数: 1879 | 10 The second approach is a bit problematic since you would have hard time
querying fields within the XML document. Integrity is not much of an
issue since you can always verify the integrity of the XML document
using triggers. This approach is really for cases where you could
potentially get thousands of columns without clear idea of the schema
and the table is sparse (some e-commerce data could be like this).
In those cases, unless you are doing research, use native XML databases
instead, such |
|
N***m
发帖数: 4460 | 11 sparse matrix has itw own way of sloving such big problems.
or you can use mpi?
1000000*
值最 |
|
s**p
发帖数: 207 | 12 i have a sparse matrix to solve, which need to change a little bit in the
future, will add 1 or 2 non-zero elements or delete 1 or 2 evertime.
i want to use CSR (compressed storage by row) to store it
one way is to implement with three vectors, val, row_ptr, col_idx
but this will give me trouble when i modified the matrix. changing only
several elements I need to re-do the whole procedure.
so I want to use a multimap. which should work very nicely.
then i have one concern, how can i deal with t |
|
T****i
发帖数: 191 | 13 看了些资料但是不得要领，我用c++生成一个sparse matrix然后想用matlab的solver求
解，因为自己的solver又慢又不准。现在我是把数据存到文件，然后matlab读取求解，
但有精度损失。后来看网上说c++可以调用matlab，但是死活看不懂，有哪位大侠能把
详细过程说说啊！感激不尽！！ |
|
s******u
发帖数: 179 | 14 在fortran 90中，我将一个稀疏矩阵存成下面的数据格式：
TYPE:: rsm !Real sparse matrix
integer:: numbers !number of nonzero value in the matrix
integer,dimension(:),pointer::rows
integer,dimension(:),pointer::columns
real ,dimension(:),pointer::values
END TYPE rsm
然后用指针读取这个矩阵
TYPE:: rsmptr
type(rsm),pointer::p
END TYPE rsmptr
每次读取都是从第一个到后一个顺序读取。后面的程序中要对这个矩阵多次重复的读取(且是在内存中)
，这样的存法，读取的效率不怎么高。我也试过把一个矩阵中的元素存成一个node的数
据结构：
type:: node
integer :: rows
integer |
|
b****p
发帖数: 10 | 15 There are lots of different storage formats for the sparse matrix. I
know there is a package called SPARSKIT. Although you do not necessarily
use the package, at least you should take a look before you develop your
own data structure. Basically you should work more on the index array
but not simply use row and column ID as the index. |
|
b**********y
发帖数: 7371 | 16 Well, as far as I know, those are not so marketable skills, but YMMV.
InterSystem Cache and Ensemble don't have much literacy in the open society.
You will need a copy of Cache or Ensemble installed to get to its local
online doc set - kinda sparse. But if you are familiar with Basic, it shud
be quick. |
|
G*****7
发帖数: 1759 | 17 4x4 -> use intel ipp smal matrix library
4k x 4k ->use intel mkl blas/linpack/sparse solvers
either way, intel ipp/mkl is free for academic use under linux and comes
with tons of examples.
子。 |
|
n**f
发帖数: 121 | 18 I thought BLAS only comes with basic operations...and not necessarily good
ones for both dense and sparse. |
|
n**f
发帖数: 121 | 19 Depending on how big/sparse the matrice are, and how frequently you have to
solve this, you may want to try LU or SVD using numerical recipe. |
|
p***o
发帖数: 1252 | 20 I am talking about large sparse matrix. |
|
c***r
发帖数: 1570 | 21 for the large sparse matrix, try petsc |
|
g*****g
发帖数: 34805 | 22 You must be kidding. This excerpt says it all.
Product group resourcing for the feature was low until 2004, and remained
sparse, leading to major overload on MSR Cambridge employees, though
ultimately the code, designs and specifications were fully transferred into
both the CLR and C#.
It's so coincidently java 1.5 was released in 04 and it had Generics.
What happens they had some academic thoughts, they talked about it maybe.
But it wasn't until java 1.5 release they copied and revised it. |
|
c*******2
发帖数: 66 | 23 Their countries are so sparsely populated. Their winter is so long.
So they get to talk to their intimate machines a lot more often ...
Their raw fish is really good. Better than sushi. |
|
y******u
发帖数: 804 | 24 贴首诗吧，引引大牛们吧
Zen of Python
Beautiful is better than ugly.
Explicit is better than implicit.
Simple is better than complex.
Complex is better than complicated.
Flat is better than nested.
Sparse is better than dense.
Readability counts.
Special cases aren't special enough to break the rules.
Although practicality beats purity.
Errors should never pass silently.
Unless explicitly silenced.
In the face of ambiguity, refuse the temptation to guess.
... 阅读全帖 |
|
|
p***o
发帖数: 1252 | 26 x又不一定稀疏，用掉24G很轻松，LU再用掉10-20G，75G也不算太离谱。
b要是只有几列你可以试试CG/GMRES之类的，有60000列的话很难比
matlab快到哪里去。 |
|
w**w
发帖数: 100 | 27 有道理，那只要 inverse A, 有什么好办法么？
好像时间主要花在 inverse A 上了？
我土，只用 python. matlab, 看看你说的库 python 有末有。
多谢大牛 |
|
p***o
发帖数: 1252 | 28 别，千万别算inv(A)。
你去python/matlab找iterative solver试，内存肯定省，时间不一定。
要不然你说说A都有啥性质，比如对称正定之类的，大家还能帮你多推荐一点。 |
|
w**w
发帖数: 100 | 29 省时间最重要。
A 没啥好的性质，就是比较稀疏而已。
多谢大牛 |
|
t*****n
发帖数: 4908 | 30 1 请忘记matlab。这个就是垃圾。
2 有无数库可以用。umfpack可能是最好的选择了。
3 算法优先考虑。再牛的库，也挡不住把数据全部放入内存。
4 要花时间编程。 |
|
w**w
发帖数: 100 | 31 呵呵。umfpack 是个好东西。我去用用看。
大牛具体说说怎么个算法好？
要花时间编程从何说起？ 难道你看出我的编程水平了？ 汗啊，我的编程真是差的很，
指点一二？
我解过 size A,b = 20000 20000 左右同样的方程，matlab 用半分钟，python (就是
那3行糟糕程序) 用1－2分钟.
没想到矩阵大点就慢这么多。。。 |
|
p***o
发帖数: 1252 | 32 matlab的稀疏矩阵缺省不就是用的umfpack当solver么 ... |
|
t*****n
发帖数: 4908 | 33 稀疏矩阵也有好多类型。比如常见的是带状稀疏。非零元素都集中在对角线上。对稀疏
矩阵，通常是只存储非零元素，并用一维数组存储。umfpack有例子。你看看就可以了。
算法方面最直接的就是分而治之。直接让计算机搞几万乘几万的矩阵是不现实的。
你现在准备做什么？有物理模型吗？ |
|
t*****n
发帖数: 4908 | 34 估计lz直接求逆矩阵了。。。。matlab我好久没摸了。 |
|
t****t
发帖数: 6806 | 35 matlab在做矩阵方面还是有它的优势, 说垃圾有点过了. 不说别的, 就说包装好这么多
库让你几乎透明使用, 就是一件挺大的工程. |
|
|
|
g****u
发帖数: 252 | 38 你这个矩阵运算matlab要多少时间就是多少时间, 不可能快很多了. |
|
p*****y
发帖数: 1049 | 39 这两天我恰好也在研究这个问题，
我知道最快的办法是conjugate gradient iterative method
线性算法 |
|
N***r
发帖数: 2539 | 40 try the open source package PETSc
it needs a bit learning though |
|
N***r
发帖数: 2539 | 41 it requires matrix A to be symmetric, positive definite and real |
|
a****l
发帖数: 8211 | 42 你是用专用的处理稀疏矩阵的库函数，还是直接把两个矩阵输入matlab里面的？5万大
小的矩阵，能有2.5G个元素，直接存在内存里面的，使用double的话就是20G的内存。 |
|
w**w
发帖数: 100 | 43 I will try it ... need figure out the matrix is symmetric and positive first
...
thanks |
|
|
W**********r
发帖数: 61 | 45 you need reorder the matrix A first to minimize fill-in elements. And
perform symbolical LDU factorization for A. Then you can use Newton-Raphson
algorithm to solve. You can use Matlab, Python, C, whaterever u want.
BTW, remember to use subroutines. |
|
N***r
发帖数: 2539 | 46 好新颖的preconditioning的方式，好像不常见啊。
我觉得你的问题是，虽然A^T仍然是稀疏矩阵，A^T A 未必仍然是稀疏矩阵吧。
如果你这个方法这么好用，这么多preconditioning的方法岂不是白瞎了。 |
|
g****t
发帖数: 31659 | 47 A'A然后共轭梯度，这个就是教科书上的方法。
还不常见？ |
|
N***r
发帖数: 2539 | 48 ahh i see
wiki上都有。不过上面说如果A是ill conditioned，那么A_T A更加ill condition，还
是需要preconditioning，所以我说A_T A并不是一个很好的preconditioning过程。 |
|
t*******e
发帖数: 1633 | 49 umfpack下载c code在matlab里compile，包里有傻瓜教程，速度非常满意 |
|
g****t
发帖数: 31659 | 50 你最后一句话是错的,至少是misleading.
例如Cholesky分解也需要这么做,但Cholesky基本上可以说是
numerical stability最强的求逆方法了. |
|
