[DRBD-user] Protocol A vs Protocol C performance issue?

Digimer lists at alteeve.ca
Tue Oct 20 20:04:12 CEST 2015

Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.


On 20/10/15 01:52 PM, Joeri Casteels wrote:
> 
>> On 20 Oct 2015, at 19:20, Digimer <lists at alteeve.ca> wrote:
>>
>> On 20/10/15 12:25 PM, Joeri Casteels wrote:
>>>
>>>> On 20 Oct 2015, at 17:47, Lionel Sausin <ls at numerigraphe.com> wrote:
>>>>
>>>> Le 20/10/2015 17:30, Joeri Casteels a écrit :
>>>>>> Protocol A an async, C is sync. So you're probably hitting the network
>>>>>> limit on C.
>>>>> I don’t think so since it has a direct 20G trunk in-between with perf i get 19.6Gbit/s on single threat. With A (async when i monitor the network it also hits 2x the network speed as C does)
>>>> "sync" means the round-trip time (down to the remote disks) is what counts, not the bandwidth.
>>> So what causes the 1/2 difference then if i read on forum’s most people don’t even see a speed difference between protocol A and C… btw both’s primary and slave node are identical hardware wise so it’s not that the disks are the limiting factor.
>>
>> It could easily be the hardware, or configuration, or any number of things.
>>
>> To diagnose, you'll want to test each node's storage on its own, test
>> the network links in isolation, etc. You need to find the location of
>> the performance issue (or confirm there is no issue) before you burn
>> time debugging higher-level applications like DRBD.
>>
> I did all the debugging on lower level before posting here, sorry i should have mentioned this but i thought that was logical :-)
> So no bottlenecks on local level and also not on network isolated level…
> 
> 
>> Protocol A says to call the write complete when it's on the local node's
>> network send buffers, so the slow-down could be on the peer's network
>> receive buffers. Try testing Protocol B. That calls the write complete
>> when it's received by the peer, but not yet committed to persistent storage.
> 
> I tried protocol B and also have the same effect as protocol C so 1/2 performance drop compared to protocol A… i played with send and receive buffers but still no change…
> I did find a performance increase today with the al-extents but that’s on all protocols so still end up with 1/2 the performance on protocol B and C.

Strikes me as something to do with the sender's network... Have you
tested in the other direction?

-- 
Digimer
Papers and Projects: https://alteeve.ca/w/
What if the cure for cancer is trapped in the mind of a person without
access to education?



More information about the drbd-user mailing list