<div><br><div class="gmail_quote">On Thu, Oct 25, 2012 at 1:14 AM, Björn Gustavsson <span dir="ltr"><<a href="mailto:bgustavsson@gmail.com" target="_blank">bgustavsson@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
On Wed, Oct 24, 2012 at 5:58 PM, Erik Pearson <<a href="mailto:erik@defunweb.com" target="_blank">erik@defunweb.com</a>> wrote:<br>
[...]<br>
<div>> And from the section on optimizations:<br>
><br>
> The basis of this optimization is that it the emulator can create bit<br>
> strings with extra uninitialized space, so if a bit string is built by<br>
> continuously appending to a binary the data does not need to be copied if<br>
> there is enough uninitialized data at the end of the bit string.<br>
><br>
> One point I don't quite understand here:<br>
><br>
> <a href="http://www.erlang.org/doc/efficiency_guide/binaryhandling.html#match_context" target="_blank">http://www.erlang.org/doc/efficiency_guide/binaryhandling.html#match_context</a><br>
><br>
> is is this function:<br>
><br>
> my_list_to_binary(List) -><br>
> my_list_to_binary(List, <<>>).<br>
><br>
> my_list_to_binary([H|T], Acc) -><br>
> my_list_to_binary(T, <<Acc/binary,H>>);<br>
> my_list_to_binary([], Acc) -><br>
> Acc.<br>
><br>
> the first iteration in my_list_binary/2 that Acc will be copied, but it will<br>
> not be copied after this. I don't know why any copying occurs at all, since<br>
> it is evident that in this case the initial binary <<>> is used solely for<br>
> the append operation in the first clause. From what I understand, it is<br>
> multiple references that cause copying to occur upon append, since the<br>
> append operation may result the movement of the binary in memory, and the<br>
> other references would otherwise become invalid. I only see one reference<br>
> threading through this function.<br>
<br>
</div>Appending to a binary is optimized by the run-time system (which is<br>
stated in the Efficiency Guide) with no help from the compiler. Therefore,<br>
the run-time system has no way of knowing that the binary created in<br>
my_list_to_binary/1 will be appended to, so it will *not* mark the empty<br>
binary as appendable and reserve extra space in it for appending.<br>
<div><br></div></blockquote><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div>
><br>
> I would think it would be allocated in my_list_to_binary/1, passed by<br>
> reference, have an initial size of 256, be appended to without any copying,<br>
> and only maybe be copied if the accumulation exceeded 256 bytes (the binary<br>
> may be moved when it is extended.) Is the implication is that the initial<br>
> blank binary <<>> is a different type of binary, say fixed-length or static,<br>
> and that the first append operation triggers the copy of that binary into a<br>
> newly allocated extendable binary (with initial size twice the original or<br>
> 256 whichever is bigger)?<br>
<br>
</div>Yes, correct. The initial empty binary is a different type of binary. It is<br>
a heap binary stored in the constant pool for the module, so the<br>
"creation" of it is extremely cheap.</blockquote><div><br></div><div>Okay, this raises, for me at least, a couple of questions:</div><div><br></div><div>- The efficiency guide mentions two storage areas for binaries, process heap and an area outside of the process heap.</div>
<div>- small binaries (up to 64 bytes) are stored on the process heap, larger ones in the shared "binary" storage.</div><div><br></div><div>- there is no mention in the eff. guide of storage of binaries in the constant pool, but if they are stored there as well, then wouldn't there just be a pointer reference to it and nothing in the heap?</div>
<div><br></div><div>- Some operations, such as copying, might create a small binary on heap</div><div>- Appending will always create a large binary, since the minimum size for an appendable binary is 256.</div><div><br></div>
<div>- Are there runtime efficiency strategies for re-using large binaries, or are they always allocated when needed? Or is allocation fast enough to offset any possible advantage of pooling?</div><div><br></div><div>- And finally, the eff. guide suggests creation of a binary by copying bytes from one to another via append, yet it is also stated that sub-binaries and match contexts are cheaper than copying. Therefore I would expect that the most efficient method of creating a new binary would be to create a sub-binary by something like part, split, or pattern matching + destructuring, as Dimitry pointed out earlier in this thread (even though my tests didn't find it to be faster.)</div>
<div><br></div><div>It may be that real world testing under load would reveal different patterns -- e.g. as excessive memory usage strains the system.</div><div><br></div><div>Thanks,</div><div>Erik.</div><div><br></div><div>
<br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> </blockquote><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<span><font color="#888888">
/Björn<br>
<br>
--<br>
Björn Gustavsson, Erlang/OTP, Ericsson AB<br>
</font></span></blockquote></div><br></div>