241 lines
12 KiB

Date: January 5, 2006
Author: Daniel Stenberg
Status of project Hiper - high performance libcurl modifications
What is Hiper
You won't find such a description in this document. See for further details.
Live Progress Info
During my work, I've posted occational updates on the curl-library mailing
list but more importantly done frequent updates of
I took time off my regular job during Decemember 2005 and the first week of
January 2006 to work on hiper full-time.
Step 1 - Measure the Existing Solution
I started full-time work on project Hiper on December 1st 2005. I began by
putting together a test application that used the existing API to allow me
to properly and with accuracy measure execution and transfer speeds when
doing a large amount of transfers.
I soon discovered that it was impossible to do any sensible measurements by
using live and actual URLs since the transfers were too unrelialble and
uncontrolled. I then enhanced the current HTTP server in the curl test suite
and made that support a large amount of transfers and some extra magic
"commands" that would make the server either just sit "idle" or "stream"
(continuously sending data in a never-ending stream). I then wrote up two
files using the curl test suite file format and by acessing the properly
formatted URLs on my localhost the HTTP server would either run "idle" or
run "stream".
Having this working, I patched libcurl to always only recv() a single byte
off the network each time, just to make sure that the time spent on reading
data is constant and never very long.
I adjusted the test application (actually called 'hiper') to create Y idle
transfers and Z stream transfers, had it run for N seconds and then quit and
produce a summary on stdout. Now I got very solid and repeatable results. I
started to run repeated tests and save the results when I ran into the
dreaded 1024 socket maximum limit.
One side of the problem is that the fd_set type only allows 1024 file
descriptors (on my Linux), which I had to solve by simply making my own type
with room for more connections and do ugly typecasts in the code. The other
side of the problem is that user applications have a limit imposed by the
system on the maximum amount of file descriptors it can have open and I had
to work around that by writing a special tool that runs setuid root that
increases the limit, downgrades to a normal user again and then run the
command line of your choice. This second approach has to be used for both
'hiper' and the test HTTP server. (You need to build the HTTP server with
CURL_SWS_FORK_ENABLED defined to have it do forks since it isn't desirable
to do so when running the normal curl tests.)
Now I could run my test program without problems. I decided to run the tests
with 1 stream connection and a varying amount of idle ones. I did 1001,
2001, 3001, 5001 and 9001 connections and measured how long select() and
curl_multi_perform() (including the curl_multi_fdset() call) would take in
average, over a period of 20 seconds. I ran each test 5-6 times and I used
the average time of all the runs.
The times in number of microseconds:
Connections multi_perform select
1001 3504 951
2001 7606 1988
3001 11045 2715
5001 16406 4024
9001 32147 8030
Test system
CPU: Athlon XP 2800
Linux: 2.6
glibc: 2.3.5
libcurl: 7.15.1
The only reason I stopped at 9001 connections is that my test machine ran
out of avaiable memory by then as I ran the test server on the same machine,
and I didn't want to risk the test result accuracy by having it start using
the swap during the tests.
It means that at 9000 connections we spend 40ms for each socket action, even
when only one socket ever have action.
With these 32000 microseconds curl_multi_perform() takes for 9000
connections, it loops 18000 laps which makes less than 2 microseconds per
lap. (Of course counting time/laps is an oversimplification, but anyway.)
Hopefully we should achieve less than 10 microseconds for each call to
curl_multi_socket() for an active connection.
The timing graph displayed on the libevent site (duplicated on the hiper
project page) suggests that libevent is pretty much fixed at 50 microseconds
(although I don't know what test box was used in their testing, we can
compare the select()-times from my tests and see that they are at least
resonably close).
Summing up, the current ~40 ms spent at 9000 connections could then possibly
be lowered to something around 60 us!
Step 2 - Implement curl_multi_socket API
Most of the design decisions and debates about this new API have already
been held on the curl-library mailing list a long time ago so I had a basic
idea on what approach to use. The main ideas of the new API are simply:
1 - The application can use whatever event system it likes as it gets info
from libcurl about what file descriptors libcurl waits for what action
on. (The previous API returns fd_sets which is very select()-centric).
2 - When the application discovers action on a single socket, it calls
libcurl and informs that there was action on this particular socket and
libcurl can then act on that socket/transfer only and not care about
any other transfers. (The previous API always had to scan through all
the existing transfers.)
The idea is that curl_multi_socket() calls a given callback with information
about what socket to wait for what action on, and the callback only gets
called if the status of that socket has changed.
In the API draft from before, we have a timeout argument on a per socket
basis and we also allowed curl_multi_socket() to pass in an 'easy handle'
instead of socket to allow libcurl to shortcut a lookup and work on the
affected easy handle right away. Both these turned out to be bad ideas.
The timeout argument was removed from the socket callback since after much
thinking I came to the conclusion that we really don't want to handle
timeouts on a per socket basis. We need it on a per transfer (easy handle)
basis and thus we can't provide it in the callbacks in a nice way. Instead,
we have to offer a curl_multi_timeout() that returns the largest amount of
time we should wait before we call the "timeout action" of libcurl, to
trigger the proper internal timeout action on the affected transfer. To get
this to work, I added a struct to each easy handle in which we store an
"expire time" (if any). The structs are then "splay sorted" so that we can
add and remove times from the linked list and yet somewhat swiftly figure
out 1 - how long time there is until the next timer expires and 2 - which
timer (handle) should we take care of now. Of course, the upside of all this
is that we get a curl_multi_timeout() that should also work with old-style
applications that use curl_multi_perform().
The easy handle argument was removed fom the curl_multi_socket() function
because having it there would require the application to do a socket to easy
handle conversion on its own. I find it very unlikely that applications
would want to do that and since libcurl would need such a lookup on its own
anyway since we didn't want to force applications to do that translation
code (it would be optional), it seemed like an unnecessary option. I also
realized that when we use underlying libraries such as c-ares (for DNS
asynch resolving) there might in fact be more than one transfer waiting for
action on the same socket and thus it makes the lookup even tricker and even
less likely to ever get done by applications. Instead I created an internal
"socket to easy handles" hash table that given a socket (file descriptor)
returns a list of easy handles that waits for some action on that socket.
To make libcurl be able to report plain sockets in the socket callback, I
had to re-organize the internals of the curl_multi_fdset() etc so that the
conversion from sockets to fd_sets for that function is only done in the
last step before the data is returned. I also had to extend c-ares to get a
function that can return plain sockets, as that library too returned only
fd_sets and that is no longer good enough. The changes done to c-ares have
been committed and are available in the c-ares CVS repository destined to be
included in the upcoming c-ares 1.3.1 release.
The 'shiper' tool is the test application I wrote that uses the new
curl_multi_socket() in its current state. It seems to be working and it uses
the API as it is documented and supposed to work. It is still using
select(), because I needed that during development (like until I had the
socket hash implemented etc) and because I haven't yet learned how to use
libevent or similar.
The hiper/shiper tools are very simple and initiates lots of connections and
have them running for the test period and then kills them all.
Since I wasn't done with the implementation until early January I haven't
had time to run very many measurements and checks, but I have done a few
runs with up to a few hundred connections (with a single active one). The
curl_multi_socket() invoke then takes 3-6 microseconds in average (using the
read-only-1-byte-at-a-time hack). If this number does increase a lot when we
add connections, it certainly matches my in my opinion very ambitious goal.
We are now below the 60 microseconds "per socket action" goal. It is
destined to be somewhat higher the more connections we have since the hash
table gets more populated and the splay tree will grow etc.
Some tests at 7000 and 9000 connections showed that the socket hash lookup
is somewhat of a bottle neck. Its current implementation may be a bit too
limiting. It simply has a fixed-size array, and on each entry in the array
it has a linked list with entries. So the hash only checks which list to
scan through. The code I had used so for used a list with merely 7 slots (as
that is what the DNS hash uses) but with 7000 connections that would make an
average of 1000 nodes in each list to run through. I upped that to 97 slots
(I believe a prime is suitable) and noticed a significant speed increase. I
need to reconsider the hash implementation or use a rather large default
value like this. At 9000 connections I was still below 10us per call.
Status Right Now
The curl_multi_socket() API is implemented according to how it is
documented. The man pages for curl_multi_socket and curl_multi_timeout are
both committed to CVS and are available online for easy browsing:
The hiper-5.patch I made available early morning January 5th, 2006 should
apply fine on a recent CVS checkout (at the time of this writing curl 7.15.1
is the latest public curl release but the hiper patch does not apply fine on
What is Left for the curl_multi_socket API
1 - More measuring with more extreme number of connections
2 - More testing with actual URLs and complete from start to end transfers.
I'm quite sure we don't set expire times all over in the code properly, so
there is bound to be some timeout bugs left.
What it really takes is for me to commit the code and to make an official
release with it so that we get people "out there" to help out testing it.
What is Left for project Hiper
1 - Add HTTP pipelining support
2 - Add a zero (or at least close to zero) copy interface
Neither of these points have been planned or detailed exactly how they will
be implemented.
Roadmap Ahead
I plan and hope to return to full-time hiper work later on this spring or
possibly summer to continue where I pause now. Of course some spare time
might also be spent until then to get us moving forward.