[Libwebsockets] trouble with external POLL array handling
"Andy Green (林安廸)"
andy at warmcat.com
Thu Jan 17 12:51:42 CET 2013
On 17/01/13 16:01, the mail apparently from Edwin van den Oetelaar included:
>> You are right a lookup table will do fine, and it blows through the
>> complexity and remaining iterations in the hash scheme. On a really low-end
>> ARM or MIPS box with only 4 or 8MB of memory though it's quite expensive to
>> reserve 128KBytes for this if he only expects one client at a time.
>> I have been thinking for a while that MAX_CLIENTS needs to become something
>> set at runtime by the app if it doesn't like the default (what should we
>> tell a distro packager to set MAX_CLIENTS to otherwise?), your scheme looks
>> like a good way to solve that and can also replace the library's internal
>> polling hash tables we discussed yesterday are also bloated, statically
>> allocated and complex. We should get a nice throughput increase there too
>> under heavy load.
>> So I will look at adapting (the second version of) this today both in and
>> out of the library and in the in-library case, runtime control over the
>> allocation for it.
>> Thanks a lot for studying it!
First I committed your changes to the test server here, after some minor
> Some thoughts I have.
> The size of the lookup table must be at least the largest number of FD
> that the whole application expects.
> So if program has 1000 files open but just 2 sockets you would still
> need at least 1002 entries in the lookup table, since fds are recycled
> by the OS.
> If a lookup table was to be used for quickly finding the fd -> wsi
> mapping, I think the memory should be allocated by the user code and
> passed to the library.
Understood... when I implemented it thismorning I realized the scope of
the "fd space" is not only smaller than your example of 30K typically
(it's 1024 on my Fedora box) but the user can control it through ulimit.
By having the server run as a "websocket" user or such he can also
control just what the server sees as the limit.
So I took the approach to get rid of MAX_CLIENTS and any other
configure-time decision and always malloc the tables to match
getdtablesize() seen at init. Then the user can forget about it and
just set ulimit as he would have to do anyway to get more file
descriptors available, and the library will adapt.
Overall that patch removed 100 lines and made things much nicer in that
area, thanks a lot for the work you did on the scheme.
> My possible use case would be to run this library on linux, for now it
> is just an idea but my goal is to write a little server that connects
> 200k devices with low traffic, low memory, low CPU, low latency using
> websockets, to create the Internet of Things.
Yes the reason I made the library was exactly with an embedded server in
mind. I'll do more debloating shortly for that case, like --without-client.
> If I can get that up and running quick I have a potential client/use-case.
> I do like python and all those fancy websocket-servers but this 'C'
> library can beat all of them. I am also mixing the poll loop with
> zero-mq and that brings all kinds of nice things.
> Maybe the handling of server/client could take into account that a
> server might have the memory for lookup tables but the client might
Even a small client will be OK with lookup tables for file descriptors
if he won't hold many open files / sockets. For now I think it's OK to
control it by ulimit... the main thing to enable smaller clients will be
--without-server that doesn't exist yet.
Actually most hours today went on finally getting rx flow control for
the server to work correctly, that's critical for server situations
where it's passing on to clients who cannot keep up with an incoming burst
More information about the Libwebsockets