ildefonso.camargo at gmail
Feb 3, 2012, 7:47 PM
Post #5 of 5
Re: Read performance goes really low while writing.
[In reply to]
It looks like an update to DRBD 8.4.1 (with update to kernel 3.x)
solved at least the issue with slow reads while writing.
Just wanted to share this with the list.
On Thu, Feb 2, 2012 at 2:34 PM, Jose Ildefonso Camargo Tolosa
<ildefonso.camargo [at] gmail> wrote:
> On Thu, Feb 2, 2012 at 11:44 AM, Brian R. Hellman <brian [at] linbit> wrote:
>> On 02/01/2012 05:16 PM, Jose Ildefonso Camargo Tolosa wrote:
>>> I'm doing some experimentation with DRBD and relatively low speed
>>> links (~30Mbps), with latency around 34ms. I have this:
>> ^^^ I think that's where your problem is. Your buffers are probably
>> filling up causing the slow down, which is also why you don't see the
>> issue while DRBD isn't connected. Even with Protocol A your buffer is
>> only so big. In situations like this we recommend people use DRBD
>> Proxy. If you're interested in giving it a try let us (LINBIT) know.
>> We use it internally to replicate data from Beaverton, OR to Vienna,
>> Austria and it works quite well.
> Yeah, maybe, but the question still remains: why writing puts reads on
> hold? why if DRBD is supposed to read locally, it doesn't allow reads
> while writing?
>>> 1. DRBD version: 8.3.7 (api:88/proto:86-91)
>>> 2. Several virtual machine images on the DRBD volume.
>>> 3. deadline scheduler by default.
>>> 4. LVM on top of DRBD, and ext4 on top of this LVM.
>>> I did this:
>>> 1. I started 3 KVM VMs (mainly idle).
>>> 2. I started to copy a big file (50GB) into DRBD partition.
>>> 3. I started to use the VMs, mainly read-only operations, with some
>>> small writes.
>>> At some point, the VMs just froze while trying to read. So, I went to
>>> physical machine, and tried a "ls" on a directory inside the DRBD
>>> volume, that I hasn't saw before (so it was not cached)... ls command
>>> also froze.... access to cached data seems to be unaffected, then, I
>>> stopped the cp process (ctrl-z), and waited for network traffic to
>>> stop (patiently watching bwm). After network (replication) traffic
>>> stopped: all unfroze.
>>> Now, my understanding is that DRBD reads are local, and thus: I can't
>>> understand why writing would freeze the reads....
>>> Any ideas?
>>> Thanks in advance,
>>> Ildefonso Camargo
>>> drbd-user mailing list
>>> drbd-user [at] lists
>> : Brian Hellman
>> : LINBIT | "Your Way to High Availability"
>> : 1-503-573-1262 x202 | 1-877-4-LINBIT
>> : Web: http://www.linbit.com
>> : Twitter: http://www.linbit.com/en/twitter
>> : Facebook: http://www.linbit.com/en/facebook
>> drbd-user mailing list
>> drbd-user [at] lists
drbd-user mailing list
drbd-user [at] lists