上一次说到了实现一个简单cache 的基本思路和想法, http://www.cnblogs.com/--00/p/erlang_ets_something_about_cache.html 在文末, 说到了判断single record 内存占用量. 这次继续说说Erlang 数据项内存的相关问题.
在Erlang efficiency_guide 文档中, 较为清楚的表述了Erlang 系统中不同数据类型的内存消耗, 在这简单贴一两个:
Small integer | 1 word On 32-bit architectures: -134217729 < i < 134217728 (28 bits) On 64-bit architectures: -576460752303423489 < i < 576460752303423488 (60 bits) |
List | 1 word + 1 word per element + the size of each element |
Atom | 1 word. Note: an atom refers into an atom table which also consumes memory. The atom text is stored once for each unique atom in this table. The atom table is not garbage-collected. |
String (is the same as a list of integers) | 1 word + 2 words per character |
从文档中,可以看出Small integer 占用了1个字节, Atom 占用1个字节, List 占用的字节主要取决于element amount 和 size of each element .
举个栗子:
["123", "234"] 占用的内存量的计算 1 + (1 + (1 + 2 * 3)) + (1 + (1 + 2 * 3)) = 17 就是17 个字节.
Tips:
注意Atom 在Erlang 系统中只占用1 个word, 这一点对于Message 有很大的帮助.
Erlang中atom数据类型能够做的唯一的运算就是比较;在erlang中模块名和方法名都是原子;Atom用来构造Tag-Message,Atom的比较时间是常量的,与Atom的长度无关(如果拿binary做tag,比较时间是线性的);Atom就是为比较而设计,除了比较运算不要把Atom用在别的运算中.
扩展阅读参见坚强的blog.
了解了Erlang 各种数据项在Erlang 系统中的内存分配规则,那么怎么才能快速的计算呢? 有没有现成的API函数, 总不能每次都手动计算一次吧?
那就首先来看看Erlang 系统所提供的各种size:
- 其中对于所有数据项都通用的有:
erlang:external_size/1
,erts_debug:size/1
,erts_debug:flat_size/1
- 适用于二进制串有:
erlang:size/1
,erlang:byte_size/1
,erlang:bit_size/1
- 适用于元组的有:
erlang:size/1
,erlang:tuple_size/1
其中,比较重要的erts_debug 两个函数:
erts_debug:size/1
和erts_debug:flat_size/1
都是不在正式文档中的函数, 可以用来计算erlang数据项在内存中所需要空间. 各种数据项的空间占用可以在这里找到: http://www.erlang.org/doc/efficiency_guide/advanced.html#id68912. 这两个函数区别在于, 在具有共享内存的数据结构中,erts_debug:size/1
只计算一次共享的数据大小, 而erts_debug:flat_size/1
则会重复计算.这是erlang源代码中的例子:
%% size(Term) %% Returns the size of Term in actual heap words. Shared subterms are %% counted once. Example: If A = [a,b], B =[A,A] then size(B) returns 8, %% while flat_size(B) returns 12.
文档中有另外一个例子: http://www.erlang.org/doc/efficiency_guide/processes.html
总的来说,
erts_debug:size/1
是erlang数据项在内存中所占用的空间大小,erts_debug:flat_size/1
是同一节点内, 跨进程移动数据项(包括ETS操作)所需要拷贝的数据大小.
OK, 先做个简单的test :
1 $ cat test_for_ets_record_flat_size.erl 2 -module(test_for_ets_record_flat_size). 3 4 -compile(export_all). 5 6 start() -> 7 A = ets:new(a, [named_table, public]), 8 D = {[{} || _ <- lists:seq(1, 100)], [self() || _ <- lists:seq(1, 10)], [{<<"1234567890">>, {}} || _ <- lists:seq(1, 1000)]}, 9 io:format(" ** data words size ~p~n", [erts_debug:flat_size(D)]), 10 io:format(" ** before insert ~p~n", [ets:info(A, memory)]), 11 ets:insert(A, D), 12 io:format(" ** after insert ~p~n", [ets:info(A, memory)]).
执行结果:
1 $ erl 2 Erlang/OTP 17 [erts-6.3] [source] [64-bit] [smp:8:8] [async-threads:10] [hipe] [kernel-poll:false] [dtrace] 3 4 Eshell V6.3 (abort with ^G) 5 1> test_for_ets_record_flat_size:start(). 6 ** data words size 10324 7 ** before insert 305 8 ** after insert 10633 9 ok
从结果来看, flat_size 的方法差了4个字节. (erts_debug:size/1 能相差 8095 个字节) 为什么?
erts_debug:size/1 只计算一次共享的数据大小, 而erts_debug:flat_size/1
则会重复计算. 但为什么erts_debug:flat_size 还是会相差4个字节呢?
带着这个疑惑去Google erlang erts_debug flat_size 找到了 Erlang-MailList :
After looking at this more I have realized the documentation of the memory information is correct as would be expected. Sorry for the noise about this. Some comment that talks about erts_debug:flat_size/1 (and erts_debug:size/1) providing the process heap size only, with an additional 1 word excluded for the register or stack storage of the top-level term would help make things clearer. This may be straight-forward for some since it makes logical sense, but I didn‘t know about these internal details and I wanted to be sure I was looking at the size correctly.
提到了erts_debug:flat_size ONLY 提供占用进程heap size .
回过头看 源代码:
Returns the size of Term in actual heap words.
事情进展到这, 有一点已经搞明白了: erts_debug:flat_size/1 只能计算Erlang Term 在进程heap 中占用的内存, 并不能计算所有的内存占用量.然后通过上面的Erlang-MailList 找到了github 上的这个开源项目: erlang_term
摘取其中关键性的一段代码:
1 byte_size_term(Term, WordSize) -> 2 DataSize = if 3 is_binary(Term) -> 4 BinarySize = erlang:byte_size(Term), 5 if 6 BinarySize > 64 -> 7 BinarySize; 8 true -> 9 % in the heap size 10 0 11 end; 12 true -> 13 0 14 end, 15 % stack/register size + heap size + data size 16 (1 + erts_debug:flat_size(Term)) * WordSize + DataSize.
从上面的代码可以看出, Erlang Term 的内存占用量应该是process heap 的内存占用量(通过erts_debug:flat_size/1 计算), stack 占用量以及共享内存占用量的总和.
好, 继续上test code :
1 $ cat test_for_ets_record.erl 2 -module(test_for_ets_record). 3 4 -compile(export_all). 5 6 start() -> 7 A = ets:new(a, [named_table, public]), 8 D = {[{} || _ <- lists:seq(1, 100)], [self() || _ <- lists:seq(1, 10)], [{<<"1234567890">>, {}} || _ <- lists:seq(1, 1000)]}, 9 io:format(" ** data words size ~p~n", [erlang_term:byte_size(D)/8]), 10 io:format(" ** before insert ~p~n", [ets:info(A, memory)]), 11 ets:insert(A, D), 12 io:format(" ** after insert ~p~n", [ets:info(A, memory)]).
测试结果:
1 $ erl -pa ./ebin -pa ./ 2 Erlang/OTP 17 [erts-6.3] [source] [64-bit] [smp:8:8] [async-threads:10] [hipe] [kernel-poll:false] [dtrace] 3 4 Eshell V6.3 (abort with ^G) 5 1> test_for_ets_record:start(). 6 ** data words size 10325.0 7 ** before insert 305 8 ** after insert 10633 9 ok
why ?? 为什么还差3个字节呢? 好吧, 只能开issues请教作者了.
So, the erlang_term module can help you manage caching, but the real situation in the Erlang VM with the many memory pools is more complex.
没辙了, 要搞清楚这3个字节在ets table 中用到了什么地方, 就需要详细了解ets 的内存管理方式.只能先暂时搁置了(待续).
总结:
1, erts_debug:flat_size/1 只计算了Erlang Term 在process heap 中的size ;
2, erlang_term is so amazing .