hadoop-common-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Jason Venner <ja...@attributor.com>
Subject Re: 0.18.1 datanode psuedo deadlock problem
Date Mon, 12 Jan 2009 19:50:39 GMT
There is no reason to do the block scans. All of the modern kernels will 
provide you notification when an file or directory is altered.

This could be readily handled with a native application that writes 
structured data to a receiver in the Datanode, or via JNA/JNI for pure 
java or mixed solutions.



Raghu Angadi wrote:
> Sagar Naik wrote:
>> Hi Raghu,
>>
>>
>> The periodic "du" and block reports thread thrash the disk. (Block 
>> Reports takes abt on an avg 21 mins )
>>
>> and I think all the datanode threads are not able to do much and freeze
>
> yes, that is the known problem we talked about in the earlier mails in 
> this thread.
>
> When you have millions of blocks, one hour for du and block report 
> intervals is too often for you. May be you could increase it to 
> something like 6 or 12 hours.
>
> It still does not fix the block report problem since DataNode does the 
> scan in-line.
>
> As I mentioned in earlier mails, we should really fix the block report 
> problem. As simple fix would scan (very slowly, unlike DU) the 
> directories in the background.
>
> Even after fixing block reports, you should be aware that excessive 
> number of block does impact the performance. No system can guarantee 
> performance when overloaded. What we want to do is to make Hadoop 
> degrade gracefully.. rather than DNs getting killed.
>
> Raghu.

Mime
View raw message