[Gluster-users] libgfapi access
Ankireddypalle Reddy
areddy at commvault.com
Tue Dec 15 20:21:20 UTC 2015
Thanks for the explanation. Valgrind profiling shows multiple memcpy's being invoked for each write through libgfapi. Is there a way to avoid these memcpy's?. Also is there a limit on the number of glfs_t* instances that can be allocated at a given point of time. I've encountered cases where if more than 8 glfs_t* instances are being allocated then glfs_init fails.
Thanks and Regards,
Ram
-----Original Message-----
From: Pranith Kumar Karampuri [mailto:pkarampu at redhat.com]
Sent: Monday, December 14, 2015 11:13 PM
To: Ankireddypalle Reddy; Vijay Bellur; gluster-users at gluster.org
Subject: Re: [Gluster-users] libgfapi access
On 12/11/2015 08:58 PM, Ankireddypalle Reddy wrote:
> Pranith,
> Thanks for checking this. Though the time taken to run was 18 seconds if you look at the time consumed in user land as well as kernel land for executing the command then it is evident that fuse took almost half the time as libgfapi. Also from the collected profiles it is evident that the average latency for the write command is less for fuse than for libgfapi. Are there any recommendations for I/O through libgfapi for disperse volumes. Is there any way to avoid the extra memcpy's that are being made when performing I/O through libgfapi.
hi Ankireddy,
Oh this is not a problem. If we use fuse, the system call 'write' from ./GlusterFuseTest will go through fuse-kernel, fuse kernel sends the write operation to glusterfs mount process which is a user process. Time taken to complete that call from then on is computed against the glusterfs mount process until it responds to the fuse-kernel, not against the ./GlusterFuseTest process. If we use gfapi, there is no system call over head, instead ./GlusterFuseTest process directly makes calls with the bricks through gfapi library. So all the time that the process spends communicating with the bricks and getting the response is counted against ./GlusterFuseTest. That is the reason you see more 'user' time.
So again, There are quite a few workloads where gfapi has proven to give better response times than fuse mounts because we avoid the context switch costs of ./GlusterFuseTest -> fuse-kernel -> glusterfs-mount -> fuse-kernel (for response)-> ./GlusterFuseTest (for response to 'write')
Hope that helps. Sorry for the delay in response, was in too many meetings yesterday.
Pranith
>
> Thanks and Regards,
> Ram
>
> -----Original Message-----
> From: Pranith Kumar Karampuri [mailto:pkarampu at redhat.com]
> Sent: Thursday, December 10, 2015 10:57 PM
> To: Ankireddypalle Reddy; Vijay Bellur; gluster-users at gluster.org
> Subject: Re: [Gluster-users] libgfapi access
>
>
>
> On 12/10/2015 07:15 PM, Ankireddypalle Reddy wrote:
>> Hi,
>> Please let me know in case you need any more details. Even for only write operations fuse seems to outperform libgfapi. Is it because of disperse volumes?. Also I noticed a lot of data loss in case I use libgfapi asyn I/O for disperse volumes.
> Fuse and gfapi seem to take same amount of time to complete the run, i.e. 18 seconds. Could you let me know what you mean by fuse outperforming gfapi?
>
> Pranith
>> Thanks and Regards,
>> Ram
>>
>> -----Original Message-----
>> From: Ankireddypalle Reddy
>> Sent: Wednesday, December 09, 2015 5:01 PM
>> To: 'Pranith Kumar Karampuri'; Vijay Bellur;
>> gluster-users at gluster.org
>> Subject: RE: [Gluster-users] libgfapi access
>>
>> Hi,
>> I upgraded my setup to gluster 3.7.3. I tested writes by performing writes through fuse and through libgfapi. Attached are the profiles generated from fuse and libgfapi. The test programs essentially writes 10000 blocks each of 128K.
>>
>> [root at santest2 Base]# time ./GlusterFuseTest /ws/glus 131072 10000 Mount path: /ws/glus Block size: 131072 Num of blocks: 10000 Will perform write test on mount path : /ws/glus Succesfully created file /ws/glus/1449697583.glfs Successfully filled file /ws/glus/1449697583.glfs Write test succeeded Write test succeeded.
>>
>> real 0m18.722s
>> user 0m3.913s
>> sys 0m1.126s
>>
>> [root at santest2 Base]# time ./GlusterLibGFApiTest dispersevol santest2
>> 24007 131072 10000 Host name: santest2
>> Volume: dispersevol
>> Port: 24007
>> Block size: 131072
>> Num of blocks: 10000
>> Will perform write test on volume: dispersevol Successfully filled file 1449697651.glfs Write test succeeded Write test succeeded.
>>
>> real 0m18.630s
>> user 0m8.804s
>> sys 0m1.870s
>>
>> Thanks and Regards,
>> Ram
>>
>>
>>
>> -----Original Message-----
>> From: Pranith Kumar Karampuri [mailto:pkarampu at redhat.com]
>> Sent: Wednesday, December 09, 2015 1:39 AM
>> To: Ankireddypalle Reddy; Vijay Bellur; gluster-users at gluster.org
>> Subject: Re: [Gluster-users] libgfapi access
>>
>>
>>
>> On 12/08/2015 08:28 PM, Ankireddypalle Reddy wrote:
>>> Vijay,
>>> We are trying to write data backed up by Commvault simpana to glusterfs volume. The data being written is around 30 GB. Two kinds of write requests happen.
>>> 1) 1MB requests
>>> 2) Small write requests of size 128 bytes. In case of libgfapi access these are cached and a single 128KB write request is made where as in case of FUSE the 128 byte write request is handled to FUSE directly.
>>>
>>> glusterfs 3.6.5 built on Aug 24 2015 10:02:43
>>>
>>> Volume Name: dispersevol
>>> Type: Disperse
>>> Volume ID: c5d6ccf8-6fec-4912-ab2e-6a7701e4c4c0
>>> Status: Started
>>> Number of Bricks: 1 x (2 + 1) = 3
>>> Transport-type: tcp
>>> Bricks:
>>> Brick1: ssdtest:/mnt/ssdfs1/brick3
>>> Brick2: sanserver2:/data/brick3
>>> Brick3: santest2:/home/brick3
>>> Options Reconfigured:
>>> performance.cache-size: 512MB
>>> performance.write-behind-window-size: 8MB
>>> performance.io-thread-count: 32
>>> performance.flush-behind: on
>> hi,
>> Things look okay. May be we can find something using profile info.
>>
>> Could you post the results of the following operations:
>> 1) gluster volume profile <volname> start
>> 2) Run the fuse workload
>> 3) gluster volume profile <volname> info > /path/to/file-1/to/send/us
>> 4) Run the libgfapi workload
>> 5)gluster volume profile <volname> info > /path/to/file-2/to/send/us
>>
>> Send both these files to us to check what are the extra fops if any that are sent over network which may be causing the delay.
>>
>> I see that you are using disperse volume. If you are going to use disperse volume for production usecases, I suggest you use 3.7.x preferably 3.7.3. We fixed a bug in releases from 3.7.4 till 3.7.6 which will be released in 3.7.7.
>>
>> Pranith
>>> Thanks and Regards,
>>> Ram
>>>
>>>
>>> -----Original Message-----
>>> From: Vijay Bellur [mailto:vbellur at redhat.com]
>>> Sent: Monday, December 07, 2015 6:13 PM
>>> To: Ankireddypalle Reddy; gluster-users at gluster.org
>>> Subject: Re: [Gluster-users] libgfapi access
>>>
>>> On 12/07/2015 10:29 AM, Ankireddypalle Reddy wrote:
>>>> Hi,
>>>>
>>>> I am trying to use libgfapi interface to access
>>>> gluster volume. What I noticed is that reads/writes to the gluster
>>>> volume through libgfapi interface are slower than FUSE. I was
>>>> expecting the contrary. Are there any recommendations/settings
>>>> suggested to be used while using libgfapi interface.
>>>>
>>> Can you please provide more details about your tests? Providing information like I/O block size, file size, throughput would be helpful.
>>>
>>> Thanks,
>>> Vijay
>>>
>>>
>>>
>>>
>>>
>>> ***************************Legal
>>> Disclaimer***************************
>>> "This communication may contain confidential and privileged material
>>> for the sole use of the intended recipient. Any unauthorized review,
>>> use or distribution by others is strictly prohibited. If you have
>>> received the message by mistake, please advise the sender by reply email and delete the message. Thank you."
>>> ********************************************************************
>>> *
>>> * _______________________________________________
>>> Gluster-users mailing list
>>> Gluster-users at gluster.org
>>> http://www.gluster.org/mailman/listinfo/gluster-users
>>
>>
>> ***************************Legal
>> Disclaimer***************************
>> "This communication may contain confidential and privileged material
>> for the sole use of the intended recipient. Any unauthorized review,
>> use or distribution by others is strictly prohibited. If you have
>> received the message by mistake, please advise the sender by reply email and delete the message. Thank you."
>> *********************************************************************
>> *
>
>
>
> ***************************Legal Disclaimer***************************
> "This communication may contain confidential and privileged material
> for the sole use of the intended recipient. Any unauthorized review,
> use or distribution by others is strictly prohibited. If you have
> received the message by mistake, please advise the sender by reply email and delete the message. Thank you."
> **********************************************************************
***************************Legal Disclaimer***************************
"This communication may contain confidential and privileged material for the
sole use of the intended recipient. Any unauthorized review, use or distribution
by others is strictly prohibited. If you have received the message by mistake,
please advise the sender by reply email and delete the message. Thank you."
**********************************************************************
More information about the Gluster-users
mailing list