Hi, Matthieu
My PEs are not I/O bound, but they do send out lots of events. My design
requires one PN to deal with tens of thousands events per second.
These screen-shots are taken when not-too-many events are sent to the PEs.
The PEC.workQueue soon reaches its max size, while CPU usage remains low.
Thread 3 to Thread 8 denote the dequeuers; the linux kernel udp buffer is
set to 80M.
I just can't figure out where the problem is; is the RAM not fast enough?
Thanks.
1) CPU Usage
[image: NewCPU.png]
2) Threads
[image: newTh.png]
3) Hot spots
[image: NewHotspot.png]
On Tue, Jan 3, 2012 at 2:21 AM, Matthieu Morel <mmorel@apache.org> wrote:
> On 12/25/11 4:23 PM, Kurtt Lin wrote:
>
>> Hello, guys.
>>
>> My S4 node is equipped with quad-core X3430 processors. What I noticed
>> is that, no matter how fast I send events to the node, the node's CPU
>> usage would not exceed 360%, even if the PEContainer.workQueue.size()
>> kept growing very fast. My question is, shouldn't the processors spare
>> more time to help PEContainer to consume the events? Why are they idle?
>> When I slowed down sending the events, I even monitored the CPU usage
>> drop to around 240%, and the PEContainer.workQueue.size() still keep
>> growing, though in a very low speed.
>>
>
> Could that be related to what your PEs are actually doing? is your
> application I/O bound? Are your PEs sending lots of events?
>
> Also, you could try to profile the JVM to get more information.
>
> Let us know!
>
> Matthieu
>
|
| Mime |
- Unnamed multipart/related (inline, None, 0 bytes)
- Unnamed multipart/alternative (inline, None, 0 bytes)
|