[PVE-User] Time synchronisation

[Adam Thompson]

On 14-08-11 11:44 AM, Joel S. | VOZELIA wrote:
> Did you get any feedback or advice? I'm interested too.
>> We are running a proxmox 3.24 on top a ubuntu 12.04 and a freeswitch
>> On the freeswitch we have to transcode voice from G711 to G729 and vice versa
>> to perform this correctly we have to have the hardware and the software in
>> perfect clock sync.
>> My question is now do you have a best practice for this or does someone have
>> experience with proxmox VE and freeswitch?

I'm running a 4-node PVE cluster, each node being 2 x 4-core (+HT) Xeon 
L5520 with 48GB RAM.  None of the nodes are anywhere near capacity, so 
I'm not contending for CPU cycles or network bandwidth right now.
On top of that, I'm running AsteriskNOW (aka FreePBX Distro) in an 
all-G.711 environment.  No analog lines or analog hardware exists except 
for two ATAs.
I am using Asterisk as a B2BUA which means it's in the voice path for 
every call; I'm not transcoding, however.
I very occasionally encounter minor glitches in voice call quality when 
the link to my upstream SIP provider is congested, but I have never 
encountered any problems with calls internally.
Latency isn't quite as low as I'd like, but it's still well within 
acceptable limits.

I have done NO tweaking yet, nor have I implemented any QoS yet, because 
my results have shown that I don't need to.
I expect to have to give that VM more CPU cycles as the cluster becomes 
more heavily loaded, and I also expect to have to implement QoS at the 
network layer... I probably won't have to do network QoS on the PVE 
hosts, they each have more than enough bandwidth to the switches right 
now to accomodate VoIP without any packet loss.

Note that the vmbr (linux bridging module) does add a measurable amount 
of latency; I haven't tried VT-d yet, because I'd rather retain the 
ability to hot-migrate the VM than reduce latency by ~0.5msec.

However, my question is: what on earth does the OP mean by "hardware and 
software in perfect clock sync"?  Unless you're running DADHi cards, 
there is no hardware to keep in sync; and running DADHi (or worse, 
zaptel) through VT-d would just be... well... bizarre.  And even in that 
case, there still isn't really anything that can get out of sync; if you 
lose sync on a T1 trunk, you lose the entire set of channels - it's 
either OK or it's not, and it has nothing to do with transcoding.

Perhaps the OP is talking about some sort of PCI co-processor that 
handles transcoding for him???

-- 
-Adam Thompson
  athompso at athompso.net