out of memory with spmd

3 次查看（过去 30 天）

fpexp 2017-12-28

1
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/374747-out-of-memory-with-spmd

评论： fpexp 2017-12-29

Hello there, I have been having troubles with the use of spmd blocks with the Matlab Parallel Computing Toolbox. My question is: assume that ind_*** are sub2index vectors and P is a very large (but far smaller than my RAM) 4d matrix, and all the rest are conformed vectors. Why does the RAM inflates enormously and eventually activate the swap mem until I get a

"The client lost connection to worker 2. This might be due to network problems, or the interactive communicating job might have errored."

if I run this:

spmd
P = PP(:,:,:,:,1);
p_star = P(ind_000).*wp_0;
p_star = p_star + P(ind_100).*wp_s;
p_star = p_star + P(ind_010).*wp_t;
p_star = p_star + P(ind_001).*wp_r;
end

whereas the following does not encounter the same problem (memory just slightly inflates)

spmd                
P = PP(:,:,:,:,1);
end
spmd
p_star = P(ind_000).*wp_0;
end
spmd
p_star = p_star + P(ind_100).*wp_s;
end
spmd
p_star = p_star + P(ind_010).*wp_t;
end
spmd
p_star = p_star + P(ind_001).*wp_r;
end

what is the difference between the two processes? Thanks a mil for the answer

2 个评论
显示无隐藏无

fpexp 2017-12-29

I want to add that PP, and the ind_***, wp_* are "local" variables, therefore I suspect that the first instruction duplicates the instrumental var all together, whereas the other instructions double copy each single var and then release the RAM at each end. Does this conjecture make any sense?

fpexp 2017-12-29