我正在尝试通过apply()
发送一个2GB的CPython只读对象(可以被腌制)来分配给分布式工作者。这最终会为进程/线程(14+ GB)消耗大量内存。
有没有办法只将对象加载到内存中并让工作者同时使用该对象?
我有2个Dask系列Source_list和Pattern_list,分别包含7百万和3百万个字符串。我正在尝试从Pattern_list(3M)中找到Source_list(7M)中的所有子字符串匹配项。
为了加快子字符串搜索,我使用pyahocorasick包从Pattern_list创建一个Cpython数据结构(一个类对象)(该对象是可选择的)。
distributed.worker - WARNING - Memory use is high but worker has no data to
store to disk. Perhaps some other process is leaking memory? Process memory:
2.85 GB -- Worker memory limit: 3.00 GB
distributed.worker - WARNING - Memory use is high but worker has no
data to store to disk. Perhaps some other process is leaking
memory?
Process memory: 14.5 GB -- Worker memory limit: 16.00 GB
distributed.nanny - WARNING - Worker exceeded 95% memory budget. Restarting
流程需要超过2.5小时的处理时间,我从未见过它完成(在取消前将其运行8小时以上)。它还消耗10 GB以上的内存Source_list.str.find_all(Pattern_list)
需要超过2.5小时。# OS = Windows 10
# RAM = 16 GB
# CPU cores = 8
# dask version 1.1.1
import dask.dataframe as dd
import ahocorasick
from dask.distributed import Client, progress
def create_ahocorasick_trie(pattern_list):
A = ahocorasick.Automaton()
for index, item in pattern_list.iteritems():
A.add_word(item,item)
A.make_automaton()
return A
if __name__ == '__main__':
client = Client(memory_limit="12GB",processes=False)
# Using Threading, because, the large_object seems to get copied in memory
# for each process when processes = True
Source_list = dd.read_parquet("source_list.parquet")
Pattern_list = dd.read_parquet("pattern_list.parquet")
# Note: 'source_list.parquet' and 'pattern_list.parquet' are generated via dask
large_object = create_ahocorasick_trie(Pattern_list)
result = Source_list.apply(lambda source_text: {large_object.iter(source_text)}, meta=(None,'O'))
# iter() is an ahocorasick Cpython method
progress(result.head(10))
client.close()
简短的回答是将它包装在一个dask.delayed调用中
big = dask.delayed(big)
df.apply(func, extra=big)
Dask会根据需要移动它并将其视为自己的数据。话虽如此,它需要存在于每个工作人员身上,因此每个工作人员的RAM应该远远超过该工作人员所占用的RAM。 (至少4倍左右)。