I am working to do some testing using Ceph in a RED and non-RED
environment.  That might be a long way off though.

Can you tell me if the OSD to OSD flows are via long-lived TCP sessions for
writes or is the session set-up/torn-down once the block is ACK'ed as
durable back to the client?






On Sun, Apr 13, 2014 at 7:00 PM, Adam Clark <[email protected]> wrote:

> Most definitely, I actually had the same thoughts after I replied.
>
> The many more flows behaviour is more likely to be kept healthy with RED
> rather than pause frames.
>
> Pause frames would halt all flows at time of congestion (while keeping TCP
> windowing unaware of the congestion issues),
>
> RED would impact a fraction of the flows and keep the overall throughput
> of the system high using TCP as the primary method for individual flow
> rates.
>
> You are most likely to have congestion into an OSD node during multiple
> clients writes, or into a client node for reads from multiple OSDs.  Using
> DCB-X may inadvertently impact other Ceph IOPS that are not directly
> experiencing congestion.
>
> Would be interesting to do some tests though.
>
> Regards
>
> Adam
>
>
>
> On Mon, Apr 14, 2014 at 12:46 AM, N. Richard Solis <[email protected]>wrote:
>
>> Adam,
>>
>> Thank you so much for replying.
>>
>> The links you posted are excellent discussions for me to reference.
>>
>> What's interesting about the difference between iSCSI and Ceph/RBD type
>> storage is the fact that under Ceph the client will make use of lots of
>> individual TCP sessions to access any particular virtual block device,
>> whereas iSCSI will use a single TCP session for every block.  That's how
>> you get into the elephant-flow vs mice-flow problem in a converged network.
>>  The (elephant) iSCSI device would compete for limited bandwidth with all
>> of the other TCP flows and you get crappy storage performance as the TCP
>> loss mechanisms kick in and drop the iSCSI TCP frames.
>>
>> If I'm thinking correctly, then Ceph wouldn't have this problem since any
>> individual block device would be spread across multiple OSDs and have
>> multiple TCP sessions connected to any particular client.  So the problem
>> of a single elephant flow really doesn't develop because you don't have
>> those large single-session TCP flows.  Am I thinking correctly?
>>
>>
>>
>>
>> On Sun, Apr 13, 2014 at 5:00 AM, Adam Clark <[email protected]>wrote:
>>
>>> Heya,
>>>   I have read a bunch of stuff regarding PFC/DCB-X and iSCSI, so it
>>> might be a similar argument with Ceph given the underlying transport is TCP.
>>> DCB-X was needed as FC has no tolerance for loss and deals with
>>> congestion by instructing senders to slow down using pause frames.
>>>
>>> From what I read, it can help, but you really need to know what you are
>>> doing, as TCP has it's own congestion avoidance mechanisms it works under
>>> most use cases.  It it probably better to randomly introduce drops in
>>> congestions rather than use a Pause Frame.
>>>
>>> Check these out:
>>> http://blog.ipspace.net/2013/07/iscsi-with-pfc.html
>>> https://blogs.cisco.com/datacenter/the-napkin-dialogues-lossless-iscsi/
>>>
>>> Cisco centric, but but applicable to other vendors also.
>>>
>>> If you are planning on running your ceph traffic alongside other traffic
>>> types, I would definitely be putting it in it's own class and guaranteeing
>>> bandwidth under times of congestion, short queues (<50ms max delay) with
>>> somewhat aggressive RED thresholds would probably suffice.
>>>
>>> Cheers
>>>
>>> Adam
>>>
>>>
>>>
>>> On Fri, Apr 11, 2014 at 11:42 PM, N. Richard Solis <[email protected]>wrote:
>>>
>>>> Guys,
>>>>
>>>> I'm new to ceph in general but I'm wondering if anyone out there is
>>>> using Ceph with any of the Data Center Bridging (DCB) technologies?  I'm
>>>> specifically thinking of DCB-X support provided by the open-lldp package.
>>>>
>>>> I'm wondering if there is any benefit to be gained by making use of the
>>>> lossless Ethernet support in a properly configured network for the OSD to
>>>> OSD traffic that would normally occur on the "cluster" network.
>>>>
>>>>
>>>> _______________________________________________
>>>> ceph-users mailing list
>>>> [email protected]
>>>> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
>>>>
>>>>
>>>
>>
>
_______________________________________________
ceph-users mailing list
[email protected]
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com

Reply via email to