Comment 1:

Labels changed: added priority-later, go1.1maybe, removed priority-triage, go1.1.

Status changed to Thinking.

Comment 2 by jgrahamc:

I would like this very much. In one project at CloudFlare I have implemented some code
to recycle []byte buffers which are used to hold web pages. These were being allocated
and released to the GC for cleanup. We now have a small package that accepts them back
on a channel and maintains a list in a goroutine of available buffers. Anyone needing a
buffer just receives on a channel and gets the next item from the list or a new buffer
if none are available.
The code also keeps track of the size of the list discarding buffers if it looks like
there are 'too many'. The idea being that we don't keep buffers that were allocated
during a burst of activity.

Comment 3:

What you are proposing is essentially to provide free() function. That's the easiest and
the most efficient way to implement your interface :)
Another option is to provide generic pool for interface{}:
type Pool struct {...}
func (p *Pool) Get() interface{}
func (p *Pool) Put(res interface{}) bool
func (p *Pool) Drain() []interface{}
func (p *Pool) SetCapacity(global, local int)
And then if the buffer sizes have some reasonable limits, then one can have, say, 16
pools for blocks of size 4k, 8k, 12k, ..., 64k, and manually choose the right one. Brad,
would it suffice dl.google.com server?
The idea behind SetCapacity(global, local int) is that you can set local=0, and then it
will be a slower pool for expensive resources.
Yes, the fast version with local cache would benefit greatly from per-m caches.

Comment 4:

I've prototyped a sync.Cache component, but now I think that sync.Pool would be a better
name. Because there are LRU/timed/key-addressed/etc caches, so Cache will be a confusing
name.

Comment 5:

I think it will be useful. Even not particularly concurrent packages like fmt/io would
benefit. And caches in http/rpc will become bottlenecks when we optimize other
bottlenecks and got more cores.

Comment 6:

> And then if the buffer sizes have some reasonable limits, then one can have,
> say, 16 pools for blocks of size 4k, 8k, 12k, ..., 64k, and manually choose 
> the right one. Brad, would it suffice dl.google.com server?
I think asking for and freeing n bytes is common enough that it should be first-class. 
We shouldn't make people create their own set of ever size class []byte.

Comment 7:

The function for asking for n bytes is already there -- make([]byte, n), the function
for freeing n bytes was removed -- runtime.Free().

Comment 8:

The problem with runtime.Free is that using it incorrectly from buggy package A affects
not-buggy package B.
With my NewBytePool proposal as implemented for dl.google.com, the byte pools are owned
by whoever made them.
But for everything needing []byte that I care about right now (io Copy buffers, ioutil
Discard buffers, and dl.google.com buffers), they're all fixed size.  So I'm fine
ignoring variably-sized []byte allocation pools and not muddling this discussion.
Pools of interface{} works for me.

Comment 9:

This bug is about sync.Cache, a cache of interchangeable fixed-size items.
The idea is that a sync.Cache might replace fmt's internal cache, and it
might be used for a cache of fixed-size buffers for io.Copy. A cache of
non-fixed-size items such as variable-length byte slices is quite a bit
more complicated.

Comment 10:

Labels changed: removed go1.1maybe.

Comment 11:

I have a preliminary version of sync.Cache based on proc-local data:
https://golang.org/cl/7686043/
And here it is applied to fmt:
https://golang.org/cl/7637047
benchmark                         old ns/op    new ns/op    delta
BenchmarkSprintfInt                     363          368   +1.38%
BenchmarkSprintfInt-2                   457          249  -45.51%
BenchmarkSprintfInt-4                   850          142  -83.29%
BenchmarkSprintfInt-8                  1220           78  -93.61%
BenchmarkSprintfInt-16                 1226           58  -95.22%
BenchmarkManyArgs                      2233         2296   +2.82%
BenchmarkManyArgs-2                    1255         1316   +4.86%
BenchmarkManyArgs-4                    1250          714  -42.88%
BenchmarkManyArgs-8                    1100          429  -61.00%
BenchmarkManyArgs-16                   1528          361  -76.37%

Comment 12:

I think it's too late to add this in Go 1.1. Too much in the runtime is still in flux.

Comment 13:

Pretty please? :) The runtime part looks relatively tiny:
https://golang.org/cl/7686043/diff/6001/src/pkg/runtime/proc.c

Comment 14:

Make all the other runtime churn stop, and then we can negotiate. :-)

Comment 15:

If you put this in, we promise to test it hard in the next few weeks.

Comment 31:

This issue was updated by revision 0f93118.

R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/43990043

Comment 32:

This issue was updated by revision 46b4ed2.

R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/37720047

Comment 33:

This issue was updated by revision 93e4a9d.

R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/44080043

Comment 34:

This issue was updated by revision b682f6d.

R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/44150043

Comment 35:

This issue was updated by revision 90e9669.

R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/44150043
»»»
TBR=golang-dev
CC=golang-codereviews
https://golang.org/cl/48190043

Comment 36:

Status changed to Fixed.

CL https://golang.org/cl/8202 mentions this issue.