On 9/20/2010 9:45 PM, Pranav Desai wrote:
> 
> This is with 2.1.2. Maybe I will try the latest SVN.

If you could that would be useful as the disk layout changed a bit when
support for 4096 sector size drives was added.

> Also
> note that I am testing it with a file based cache. Do you think there
> could be something there ?

Your file size is relatively large, so if your number of clients is large
and your file based cache is small you could be running into wrap-around 
issues...
where a slow client can't write the whole object before the disk wraps.

How big of a file cache are you running ?  How many clients ?  I am using raw
disk and at 10 clients (debug build) it I am doing between 5 and 3 ops/sec or
between 45 and 75 MB/sec with a total latency of around 3 seconds but probably
spiking to more like 5+... which means I would require a cache size of at least
1GB probably a lot more to be safe (because of random latency spikes).

The code *should* recognize wrap-around, but because it is typically not a 
problem
in adequately sized production systems it might not be reported.

> 
> 
>> If can provide access to a gdb session with the crash I might be able to 
>> figure
>> out what is going on...
>>
> 
> I was short on time so couldn't debug it any further ... hopefully I
> can get it done tomorrow ...
> 
> -- Pranav
> 
>> john
>>
>>
>> On 9/20/2010 6:20 PM, Pranav Desai wrote:
>>> On Thu, Sep 16, 2010 at 3:30 PM, Pranav Desai <[email protected]> 
>>> wrote:
>>>> On Thu, Sep 16, 2010 at 2:03 PM, Leif Hedstrom <[email protected]> wrote:
>>>>> On 09/16/2010 02:45 PM, Pranav Desai wrote:
>>>>>
>>>>> On Thu, Sep 16, 2010 at 12:33 PM, Leif Hedstrom <[email protected]> wrote:
>>>>>
>>>>>  On 09/16/2010 01:26 PM, Pranav Desai wrote:
>>>>>
>>>>> Hi!
>>>>>
>>>>> I am running a load test with some video files to see . I am using
>>>>> curl-loader to generate the load. I have modified it to add a random
>>>>> number to the URLs before sending so I can test with a single URL and
>>>>> still stress the cache. The webserver is a lighttpd server with
>>>>> rewrite rules to translate the random strings back to a common URL.
>>>>> The URL is essentially a 15MB video file. I can provide more details
>>>>> on the setup if needed.
>>>>>
>>>>> Ok, I've created https://issues.apache.org/jira/browse/TS-441   with this
>>>>> information. If you can find a core file (or, run traffic_server under 
>>>>> gdb),
>>>>> and get a stack trace, that would be very helpful. Also, when it crashes,
>>>>> you might get a stack trace in /var/log/messages and/or one of the log 
>>>>> files
>>>>> in the .../var/log/trafficserver  directory.
>>>>>
>>>>>
>>>
>>> I got the stack trace. I have updated the bug with the trace, but here it 
>>> is.
>>>
>>> FATAL: HTTP.cc:1526: failed assert `!"unknown m_polarity"`
>>> ./bin/traffic_server - STACK TRACE:
>>> ./bin/traffic_server(ink_fatal+0x86)[0x6f2056]
>>> ./bin/traffic_server(_ink_assert+0x81)[0x6f0d61]
>>> ./bin/traffic_server(_ZN11HTTPHdrImpl9unmarshalEl+0x35)[0x5ea385]
>>> ./bin/traffic_server(_ZN7HdrHeap9unmarshalEiiPP14HdrHeapObjImplP11RefCountObj+0x146)[0x5df926]
>>> ./bin/traffic_server(_ZN8HTTPInfo9unmarshalEPciP11RefCountObj+0xc5)[0x5ea205]
>>> ./bin/traffic_server(_ZN7CacheVC14handleReadDoneEiP5Event+0x750)[0x6648a0]
>>> ./bin/traffic_server(_ZN19AIOCallbackInternal11io_completeEiPv+0x26)[0x66ce26]
>>> ./bin/traffic_server(_ZN7EThread13process_eventEP5Eventi+0x22f)[0x6e7c0f]
>>> ./bin/traffic_server(_ZN7EThread7executeEv+0x1aa)[0x6e810a]
>>> ./bin/traffic_server[0x6e76da]
>>> /lib64/libpthread.so.0[0x7f4a565e52e7]
>>> /lib64/libc.so.6(clone+0x6d)[0x32a18ce3bd]
>>>
>>>
>>> I will try to dig deeper, but if have any ideas or suggestions I can
>>> try those out.
>>>
>>> Thanks
>>> -- Pranav
>>
>>

Reply via email to