给大家推荐一个感觉不错的组合

hmaimi

回复 45# wpf999
http://scalemp.com/

金鹏

回复 46# hmaimi

hsl1993422的小马甲

zxj-iq180

回复 33# 金鹏


    金版，四路g34主板多少钱？

ocw

回复 48# zxj-iq180

美金七百多, 但也有六百多, 運費要一百美金左右

zxj-iq180

回复 49# ocw


   不收关税的么？我在淘宝上就找到1块tyan的4路g34主板，要9000多rmb，太黑了。

金鹏

回复 48# zxj-iq180


    大概740-850美金，国内的贵

cuda

淘宝上超微的比tyan稍便宜一些，最低6500左右，关键字搜h8qgi。

zxj-iq180

回复 52# cuda


    搜了一下超微的这款，卖家信誉没有超过两钻的，不保险

wpf999

回复 42# cuda


    看了下GROMACS 4.5的源码，现在使用了一个叫 thread_mpi 的库，不需要MPICH了

1. /*
   
2.    thread_mpi is a cross-platform threading library for applications in
   
3.    high-performance computing. It supports:
   
4. 
   
5.    - Cross-platform thread primitives (thread creation, mutexes, spinlocks,
   
6.      barriers, thread-local storage, etc.).
   
7.    - Cross-platform atomic operations (compare-and-swap, add-return, etc) for
   
8.      safe lock-free synchronization.
   
9.    - An implementation of (currently, much of) MPI, either as a drop-in
   
10.      replacement, or for use in conjunction with a networked MPI      
    
11.      implementation.
    
12.    - Shared-memory allocation and memory management (planned, as of now).
    
13.    - Basic lock-free data structures (planned, as of now).
    
14. 
    
15.    Because it can be used as a drop-in replacement for MPI, existing codes
    
16.    using MPI can start using thread_mpi without major changes in the
    
17.    source code, assuming -- and this is a big assumption -- that the code
    
18.    is thread-safe.
    
19. 
    
20.    Alternatively, networked MPI calls can be used in conjunction with   
    
21.    thread_mpi calls (simply by using     
    
22.     "#include <thread_mpi.h>"
    
23.    instead of
    
24.     "#include <tmpi.h>"
    
25.    and pre-fixing all thread_mpi MPI-like calls with tMPI instead of MPI.
    
26. 
    
27.    The availability of both MPI calls and shared-memory constructs makes it
    
28.    possible to transition (relatively) seamlessly from an MPI-style code
    
29.    to code that's optimal on multicore CPUs.
    
30. 
    
31.    Although MPI-style message passing isn't neccesarily optimal for
    
32.    performance on shared-memory systems, the MPI communicator concept and
    
33.    its emphasis on collective operations makes sense even when computing on
    
34.    one machine with multiple cores. The communicator forms the basis for
    
35.    the shared-memory allocation and lock-free data structure implementations
    
36.    in thread_mpi.
    
37. 
    
38.    Although usable as a stand-alone library, thread_mpi is designed to
    
39.    be incorporated in the code tree, eliminating any external build
    
40.    requirements. The BSD-style license that this library is distributed
    
41.    with reflects this.
    
42. 
    
43.    The atomic operations (such as compare-and-swap) are supported on:
    
44.    - gcc on x86, x86_64, PowerPC and Itanium.
    
45.    - Intel compilers on x86, x86_64 and Itanium.
    
46.    - xlc on PowerPC.
    
47.    - (partial) HP/UX compilers on Itanium.
    
48. */
    
49. 
    
50. /** \file
    
51. *
    
52. * \brief Convenience header file for non-MPI compatibility.
    
53. *
    
54. * This file includes the tMPI header file thread_mpi/tmpi.h, as well
    
55. * as thread_mpi/threads.h and thread_mpi/atomic.h header files. If you'd
    
56. * like to use the components individually, include the relevant header
    
57. * files directly.
    
58. */
    
59. 
    
60. #include "thread_mpi/atomic.h"
    
61. #include "thread_mpi/threads.h"
    
62. #include "thread_mpi/barrier.h"
    
63. #include "thread_mpi/event.h"
    
64. #include "thread_mpi/tmpi.h"
    
65. #include "thread_mpi/collective.h"
    
66. #include "thread_mpi/hwinfo.h"
    
67. 
    
68. 
    

复制代码

cuda

多谢提供信息。
可惜fah的可执行文件里面符号定位信息都被strip掉了，否则也许可以利用一下。

回复 54# wpf999

cuda

搜关键字h8qg6还可以找到一些，但最低要7500，不知有没有两钻的。

    搜了一下超微的这款，卖家信誉没有超过两钻的，不保险
zxj-iq180 发表于 2012-2-1 23:16

iwait

有类似的intel平台么？