datastore readsize and buffer chunk

Question

Peng Li 2020-5-12

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/524896-datastore-readsize-and-buffer-chunk

评论： Peng Li 2020-5-18

I am working with a single csv file of over 10G size. I'm trying to use mapreduce to perform some analyses. The program works as expected but I'd like to speed a bit by increasing readsize.

Currently, it only passes about 1500 rows each time, although by default it is 20000 rows. I have in total over 500k rows so this creates a lot of passes that significantly slow the process.

After looking into the tabulartextdatastore, it seems that the buffer chunk that has been hardcoded to be 32MB that limits this.

So my questions are:

1) What the consideration is that MATLAB applies this 32MB buffer size

2) Whether there is a good way to manipulate this as I'd like to decrease the number of passes while increase the size within each pass (I have about 400G RAM).

2 个评论
显示无隐藏无

Peng Li 2020-5-12

up :-)

Peng Li 2020-5-13

Any suggestions/comments would be very appreciated!

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Rashed Mohammed 2020-5-18

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/524896-datastore-readsize-and-buffer-chunk#answer_433066

Hi Peng Li,

The “ReadSize” name value pair of “tabularTextDatastore” specifies the number of rows to read at most. However, it is bound by the chunk size depending on the data to efficiently manage the datastore. In your case, I would suggest you to look into partitioning the datastore and read the data in parallel. Here is a link to go through.

Hope this helps!

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Peng Li 2020-5-18

Thanks for your answer, Rashed Mohammed!

I've tried datastore partitioning as well using numlabs. Within each spmd pass I used readall. A certain process after loading data requires even more space I think (categorical or replace to replace old codings within each column to human readable meaningful phrases). It made my linux machine unresponsive and many process been killed ... So I guess I should partition it into more chunks instead of into numlabs? Or within the spmd I should use read to let it automatically load whatever size it can handle?

So far I've tried another solution by using transform a datastore. It worked well too. But I'm not sure if this can be made in paralelle as it seems to me that transform only happens when I do some process on the transformed datastore. I basically created a transform function in which each column in the old datastore is being "decoded", and after that I used writeall with parameter 'UseParalelle', 'true'. Will this enable a paralelle process to transform datastore as well?

请先登录，再进行评论。