You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@activemq.apache.org by ActiveMQ Investigation <aj...@winterflood.com> on 2016/06/14 07:06:20 UTC
Active MQ Performance
Hello,
I am doing some raw performance tests on ActiveMQ to replace SonicMQ (JMS
based) on C++ side. We have CMS using OpenWire.
My test has got five Non Durable subscribers on a Topic. When I send 25000
messages, following happens:
1. Persistent - It takes 3 times (27sec) more than what it takes on SonicMQ
(9sec). I need to find out what's causing this delay and how to improve it.
I understand that publisher will wait until broker comes back to ack
messages. I am using Auto_ACK. I have looked at SonicMQ settings for IO
Buffer and socket buffer and do the same for ActiveMQ but no success.
2. Non Persistent - This one is faster as fire and forget but messages are
sometime 2 seconds in flight on subscriber which is unacceptable to our use
case. I tried setting prefetch etc thinking it's slow consumer but no
success.
Hardware is same for both. ActiveMQ install is default installation. I am
going through performance tuning guide mentioned in the forum but any quick
clues / areas to look at will be appreciated.
Sorry if my query is too subjective but I want to find out where ActiveMQ is
taking time and why. I have done raw tests just using stomp on ActiveMQ (No
CPP involved) and it was way faster as compared to Sonic. So I suspect, I
need to tweak some settings which may be causing it.
Thanks
AJ
--
View this message in context: http://activemq.2283324.n4.nabble.com/Active-MQ-Performance-tp4712952.html
Sent from the ActiveMQ - Dev mailing list archive at Nabble.com.
Re: Active MQ Performance
Posted by ActiveMQ Investigation <aj...@winterflood.com>.
Out of two issues, I am able to resolve first one (Queue taking considerable
time on persistent messages). Topic was fast because by default,
concurrentStoreAndDispatchTopics is false for Topic. For Queue, it's
enabled. But main difference was enableJournalDiskSyncs on each message.
From SAR command, I found that IO wait time was quite high. Sometime 30%. I
switched off enableJournalDiskSyncs and concurrentStoreAndDispatchQueues and
it worked.
I think it's not ideal and makes it less reliable when broker kills but it
works until I find an alternative to wait io.
Still looking into for seconding issue.
Thanks
Ajay Kumar
--
View this message in context: http://activemq.2283324.n4.nabble.com/Active-MQ-Performance-tp4712952p4713145.html
Sent from the ActiveMQ - Dev mailing list archive at Nabble.com.
Re: Active MQ Performance
Posted by ActiveMQ Investigation <aj...@winterflood.com>.
Hello Matt,
Thanks for your message. I will do these settings and see what happens. I
have attached links to code and results.
My test show these two problems:
- ActiveMQ Queue takes more time in Persistent Mode (Q28,Q31 rows in excel
sheet) as compared to Sonic.
- ActiveMQ Non Persistent Per Message Max time is very high. (Example: U59,
U56). It looks consumer is not fast enough to process them.
Overall, Maximum Time taken by any message in flight between producer and
sender in ActiveMQ is quote high.
I have copied the code to IDEONE editor (it's just a copy and IDEONE don't
have necessary libs) to share if it helps:
Producer: http://ideone.com/4nmARP
Consumer: http://ideone.com/QnFHYk
Results:
https://docs.google.com/spreadsheets/d/1hHt7EsgwGSxHR11edunWSuzzTbc88Wch8lFkBXmPxUI/edit?usp=sharing
If you can't access code on IDEONE:
Producer -
https://docs.google.com/document/d/1RbGUKYUbWUk8LcxqbCwf53_lNEfhnxfDz3TPqSNT2Dg/edit?usp=sharing
Consumer -
https://docs.google.com/document/d/1jW5qPG40yUTBaHezEpjKrwBFZriL2bpxHZcM0U_2Gfk/edit?usp=sharing
Sorry if above link passage is not as per forum protocol. I am new to the
forum.
Thanks
AJ
--
View this message in context: http://activemq.2283324.n4.nabble.com/Active-MQ-Performance-tp4712952p4713058.html
Sent from the ActiveMQ - Dev mailing list archive at Nabble.com.
Re: Active MQ Performance
Posted by Matt Pavlovich <ma...@gmail.com>.
Hi-
How big are the messages? Can you post a link to the C++ code. There may
be some gotchas in how its sent or recv'd.
It sounds like you have confirmed that you do not have any network or
disk performance issues, since SonicMQ is getting reasonable numbers on
the same server.
Out-of-the-box ActiveMQ is configured to ensure you don't run out of
memory or lose messages, so it is setup to work for very low memory
footprints.
As a starting point adjust the memoryUsage setting here:
<systemUsage>
<systemUsage>
<memoryUsage>
<memoryUsage percentOfJvmHeap="70" />
</memoryUsage>
<storeUsage>
<storeUsage limit="100 gb"/>
</storeUsage>
<tempUsage>
<tempUsage limit="50 gb"/>
</tempUsage>
</systemUsage>
</systemUsage>
Flip the openwire to nio as well:
<transportConnector name="openwire" uri="nio://0.0.0.0:0"/>
On 6/14/16 7:56 AM, ActiveMQ Investigation wrote:
> Hello Celebert,
>
> I am using apache-activemq-5.12.1.
>
> I found out the issue. My subscribers were on same machine as ActiveMQ
> install. While 5 subscribers, consuming messages on same machine as install,
> it was quite intensive. I moved subscribers to a separate machine than
> install and it took 9 seconds to send messages and almost similar to
> receive. This is about persistent messages.
>
> For non persistent, in new setup, it takes 1.7 seconds which is acceptable.
> However, consumer (gets all msgs in 6 seconds) in is showing messages
> waiting for 1+ seconds.
>
>
> For SonicMQ, time taken is consistent irrespective of consumer on machine
> where sonic install is there.
>
>
> I feel that this 9 second is still high for sending 25000 msgs. It can do
> better. Also, can this CPU and memory intensiveness improved.
>
> I will go through Artemis. Sorry for naive query but how different Artemis
> is to ActiveMQ. I haven't explored that front.
>
> Thanks
> AJ
>
>
>
>
>
>
>
>
>
> Thanks
> AJ
>
>
>
> --
> View this message in context: http://activemq.2283324.n4.nabble.com/Active-MQ-Performance-tp4712952p4712960.html
> Sent from the ActiveMQ - Dev mailing list archive at Nabble.com.
Re: Active MQ Performance
Posted by Clebert Suconic <cl...@gmail.com>.
Artemis is an initiative to make a new broker with similar features to
ActiveMQ5 which is getting closer and closer each day.
I would say It's worth giving it a try. 1.3.0 just passed the voting,
you should see the announcement soon for the release.
On Tue, Jun 14, 2016 at 8:56 AM, ActiveMQ Investigation
<aj...@winterflood.com> wrote:
> Hello Celebert,
>
> I am using apache-activemq-5.12.1.
>
> I found out the issue. My subscribers were on same machine as ActiveMQ
> install. While 5 subscribers, consuming messages on same machine as install,
> it was quite intensive. I moved subscribers to a separate machine than
> install and it took 9 seconds to send messages and almost similar to
> receive. This is about persistent messages.
>
> For non persistent, in new setup, it takes 1.7 seconds which is acceptable.
> However, consumer (gets all msgs in 6 seconds) in is showing messages
> waiting for 1+ seconds.
>
>
> For SonicMQ, time taken is consistent irrespective of consumer on machine
> where sonic install is there.
>
>
> I feel that this 9 second is still high for sending 25000 msgs. It can do
> better. Also, can this CPU and memory intensiveness improved.
>
> I will go through Artemis. Sorry for naive query but how different Artemis
> is to ActiveMQ. I haven't explored that front.
>
> Thanks
> AJ
>
>
>
>
>
>
>
>
>
> Thanks
> AJ
>
>
>
> --
> View this message in context: http://activemq.2283324.n4.nabble.com/Active-MQ-Performance-tp4712952p4712960.html
> Sent from the ActiveMQ - Dev mailing list archive at Nabble.com.
--
Clebert Suconic
Re: Active MQ Performance
Posted by ActiveMQ Investigation <aj...@winterflood.com>.
Hello Celebert,
I am using apache-activemq-5.12.1.
I found out the issue. My subscribers were on same machine as ActiveMQ
install. While 5 subscribers, consuming messages on same machine as install,
it was quite intensive. I moved subscribers to a separate machine than
install and it took 9 seconds to send messages and almost similar to
receive. This is about persistent messages.
For non persistent, in new setup, it takes 1.7 seconds which is acceptable.
However, consumer (gets all msgs in 6 seconds) in is showing messages
waiting for 1+ seconds.
For SonicMQ, time taken is consistent irrespective of consumer on machine
where sonic install is there.
I feel that this 9 second is still high for sending 25000 msgs. It can do
better. Also, can this CPU and memory intensiveness improved.
I will go through Artemis. Sorry for naive query but how different Artemis
is to ActiveMQ. I haven't explored that front.
Thanks
AJ
Thanks
AJ
--
View this message in context: http://activemq.2283324.n4.nabble.com/Active-MQ-Performance-tp4712952p4712960.html
Sent from the ActiveMQ - Dev mailing list archive at Nabble.com.
Re: Active MQ Performance
Posted by Clebert Suconic <cl...@gmail.com>.
What version ?
Can you try Artemis ?
On Tuesday, June 14, 2016, ActiveMQ Investigation <
ajay.kumar@winterflood.com> wrote:
> Hello,
>
> I am doing some raw performance tests on ActiveMQ to replace SonicMQ (JMS
> based) on C++ side. We have CMS using OpenWire.
> My test has got five Non Durable subscribers on a Topic. When I send 25000
> messages, following happens:
>
> 1. Persistent - It takes 3 times (27sec) more than what it takes on SonicMQ
> (9sec). I need to find out what's causing this delay and how to improve it.
> I understand that publisher will wait until broker comes back to ack
> messages. I am using Auto_ACK. I have looked at SonicMQ settings for IO
> Buffer and socket buffer and do the same for ActiveMQ but no success.
>
>
> 2. Non Persistent - This one is faster as fire and forget but messages are
> sometime 2 seconds in flight on subscriber which is unacceptable to our use
> case. I tried setting prefetch etc thinking it's slow consumer but no
> success.
>
> Hardware is same for both. ActiveMQ install is default installation. I am
> going through performance tuning guide mentioned in the forum but any quick
> clues / areas to look at will be appreciated.
>
> Sorry if my query is too subjective but I want to find out where ActiveMQ
> is
> taking time and why. I have done raw tests just using stomp on ActiveMQ (No
> CPP involved) and it was way faster as compared to Sonic. So I suspect, I
> need to tweak some settings which may be causing it.
>
>
> Thanks
> AJ
>
>
>
>
> --
> View this message in context:
> http://activemq.2283324.n4.nabble.com/Active-MQ-Performance-tp4712952.html
> Sent from the ActiveMQ - Dev mailing list archive at Nabble.com.
>
--
Clebert Suconic