Re: [systemd-devel] sd-bus vs gdbus on dbus-daemon
On Thu, Apr 30, 2015 at 11:13 AM, Umut Tezduyar Lindskog u...@tezduyar.com wrote: Hi Simon, On Wed, Apr 29, 2015 at 5:34 PM, Simon McVittie simon.mcvit...@collabora.co.uk wrote: On 29/04/15 15:08, Umut Tezduyar Lindskog wrote: We [1] have noticed that there could be up to %50 performance gain on using sd-bus over gdbus on dbus-daemon. ... gdbus.c - g_dbus_proxy_new_for_bus_sync() - 50 x g_dbus_proxy_call_sync() sdbus.c - sd_bus_open_system() - 50 x sd_bus_get_property() If you want to compare apples with apples, I suggest using the lower-level g_bus_get() and g_dbus_connection_call[_sync]() instead of GDBusProxy. The design and priorities of sd-bus and GDBus are not really very similar, but GDBusProxy is certainly not the closest equivalent you can get. Thanks for the tip. I will re-run the measurements with g_dbus_connection_call_sync(g_bus_get_sync(),...). Just considering the traffic in the dbus, I do believe we have compared apple to apple. But if you believe we might get even more performance with raw API calls, then that is fantastic news! Unfortunately I didn't get any better result with g_dbus_connection_call_sync. I have updated the link (https://drive.google.com/open?id=1O3FEnpdZ2auisYalKT6qJTiNV9cfDxya_qisbpGj3MQauthuser=0) with the source code if you would like to double check my work. Umut Also, if your application profile is such that (a) D-Bus is a significant factor in performance, and (b) sending 50 synchronous D-Bus messages per connection is anywhere near realistic, then you are probably not using D-Bus the way it is designed to be used. Measurement we have done was not about if dbus is the ipc we want or not. It was about comparing the performances of two libraries. It wouldn't have been fair to compare sending only 1 message on the dbus due to the performance penalty of creating a worker thread with gdbus. But it really didn't matter. 1, 2, 10, 50, in all cases gdbus (GDBusProxy) couldn't performed as good as sd-bus. See also http://permalink.gmane.org/gmane.comp.freedesktop.dbus/13663. -- Simon McVittie Collabora Ltd. http://www.collabora.com/ ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] sd-bus vs gdbus on dbus-daemon
Hi Greg, On Wed, Apr 29, 2015 at 5:49 PM, Greg KH gre...@linuxfoundation.org wrote: On Wed, Apr 29, 2015 at 04:08:50PM +0200, Umut Tezduyar Lindskog wrote: Hi, We [1] have noticed that there could be up to %50 performance gain on using sd-bus over gdbus on dbus-daemon. For this reason, we have high interest in using sd-bus. What are the plans in terms of making sd-bus API public? Details of the test [2]: gdbus.c - g_dbus_proxy_new_for_bus_sync() - 50 x g_dbus_proxy_call_sync() sdbus.c - sd_bus_open_system() - 50 x sd_bus_get_property() Two applications are run with ltrace, perf stat -e cycles, time and the results are compared. I'll echo Simon's statement here, making a call to g_dbus_proxy_call_sync() seems like an odd thing to test. Is this really how your application wants to work? Is it the normal call path that you need optimized? How many messages do you normally send, or want to send, and how big of the data blob are you wanting to transmit/receive here? We have variety of dbus clients sending small/big slow/fast sync/async messages. But we have not focused on dbus's performance in the experiment. We wanted to see the efficiency of 2 user space libraries when it comes to a very simple use case (synchronously retrieving a property). I ask as I'm trying to find how people would like to use D-Bus, if the existing dbus-daemon were sped up in various ways. Both of these traces show that userspace is sitting around for most of the time, I don't see a whole lot of actual CPU usage happening, do you? Could you please explain how did you come up with that conclusion? Granted not the top 10 calls are in g... libraries but there are many entries that program has spent time in g... library. Also the pthread. Umut thanks, greg k-h ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] sd-bus vs gdbus on dbus-daemon
On Wed, Apr 29, 2015 at 04:08:50PM +0200, Umut Tezduyar Lindskog wrote: Hi, We [1] have noticed that there could be up to %50 performance gain on using sd-bus over gdbus on dbus-daemon. For this reason, we have high interest in using sd-bus. What are the plans in terms of making sd-bus API public? Details of the test [2]: gdbus.c - g_dbus_proxy_new_for_bus_sync() - 50 x g_dbus_proxy_call_sync() sdbus.c - sd_bus_open_system() - 50 x sd_bus_get_property() Two applications are run with ltrace, perf stat -e cycles, time and the results are compared. I'll echo Simon's statement here, making a call to g_dbus_proxy_call_sync() seems like an odd thing to test. Is this really how your application wants to work? Is it the normal call path that you need optimized? How many messages do you normally send, or want to send, and how big of the data blob are you wanting to transmit/receive here? I ask as I'm trying to find how people would like to use D-Bus, if the existing dbus-daemon were sped up in various ways. Both of these traces show that userspace is sitting around for most of the time, I don't see a whole lot of actual CPU usage happening, do you? thanks, greg k-h ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] sd-bus vs gdbus on dbus-daemon
On 29/04/15 16:49, Greg KH wrote: I'll echo Simon's statement here, making a call to g_dbus_proxy_call_sync() seems like an odd thing to test. Is this really how your application wants to work? Is it the normal call path that you need optimized? To be completely clear here: I am not saying that GDBusProxy is an unusual thing for an application to be using - far from it, it is a convenient high-level API that can cover a lot of use-cases, and if I wanted to write code to (say) manipulate systemd's D-Bus interfaces, it's what I'd use. What I *am* saying is that GDBusProxy is nowhere near the performance end of the performance/convenience trade-off, either in general, among D-Bus implementations specifically, or even within GDBus - which makes it an odd thing to benchmark. If it shows up as a significant impact in your profiling, then that might well indicate that you're using the wrong tool for the job. An analogy: suppose you told me you were interested in switching from some Python code that constructs MIME emails to some C code that does the same, because you need to send patches to systemd-devel faster. Instead of worrying about how efficiently-implemented those MIME modules were, I'd be asking you why (and whether) you are sending enough patches that it makes any practical difference, and if you are, whether pushing them to a git repository might make more sense :-) -- Simon McVittie Collabora Ltd. http://www.collabora.com/ ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
[systemd-devel] sd-bus vs gdbus on dbus-daemon
Hi, We [1] have noticed that there could be up to %50 performance gain on using sd-bus over gdbus on dbus-daemon. For this reason, we have high interest in using sd-bus. What are the plans in terms of making sd-bus API public? Details of the test [2]: gdbus.c - g_dbus_proxy_new_for_bus_sync() - 50 x g_dbus_proxy_call_sync() sdbus.c - sd_bus_open_system() - 50 x sd_bus_get_property() Two applications are run with ltrace, perf stat -e cycles, time and the results are compared. [1] - Most of the work is done by Jonatan Nilsson, I have added minor points. [2] - Full implementation details - https://drive.google.com/open?id=1O3FEnpdZ2auisYalKT6qJTiNV9cfDxya_qisbpGj3MQauthuser=0 ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] sd-bus vs gdbus on dbus-daemon
On 29/04/15 15:08, Umut Tezduyar Lindskog wrote: We [1] have noticed that there could be up to %50 performance gain on using sd-bus over gdbus on dbus-daemon. ... gdbus.c - g_dbus_proxy_new_for_bus_sync() - 50 x g_dbus_proxy_call_sync() sdbus.c - sd_bus_open_system() - 50 x sd_bus_get_property() If you want to compare apples with apples, I suggest using the lower-level g_bus_get() and g_dbus_connection_call[_sync]() instead of GDBusProxy. The design and priorities of sd-bus and GDBus are not really very similar, but GDBusProxy is certainly not the closest equivalent you can get. Also, if your application profile is such that (a) D-Bus is a significant factor in performance, and (b) sending 50 synchronous D-Bus messages per connection is anywhere near realistic, then you are probably not using D-Bus the way it is designed to be used. See also http://permalink.gmane.org/gmane.comp.freedesktop.dbus/13663. -- Simon McVittie Collabora Ltd. http://www.collabora.com/ ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
