Building New Cluster 500 Agents Need Hardware Advice.

[AkkerKid]

Hi all,
I've been tasked with building a Vicidial Cluster to support 500 outbound agents dialing with 10+ to 1 ratios.
I've learned quite a bit about hardware recommendations from this board and from the manuals but hardware has changed.

I'm leaning towards building out the following:
12 Dialers:
SuperMicro 12 Node 3U "MicroCloud" SYS-5038ML-H12TRF
Single 4-core E3-1231 3.4GHz (per node)
16GB DDR3 ECC (per node)
RAID 1? SSDs for recordings (per node)
Maybe another pair for boot (per node)

4-6 WebServers:
SuperMicro 8 Node 4U SYS-F617R2-F72+
Dual Xeon Whatevers
RAID 1 SSDs

2 Master-master MySQL Servers:
SuperMicro 8 Node 4U SYS-F617R2-F72+
Dual Xeon Whatevers
Pair 200+GB SSD for Boot
Single PCI SSD for MySQL

I have some questions about "myths" I've heard and want to migrate high-density server hardware.

1) "Timing Cards"... I'll be running SIP end-to-end. No faxes. No PSTN, No T/E1. Do I buy a Digium Card just to appease the codec Gods or is there a real reason to put one of these into each dialer?

2) Intel or AMD? More cores or more clock speed? Keep in mind I'll need to put about 50 users on each dialer (if that's even possible on a single socket).

3) When clustering, should I put every agent on every dialer or have agents SIP only into 1 dialer?

4) Is Gigabit ethernet enough? I've heard of system designers using 10Gb-E between servers, especially SQL. Is that warranted?

5) Should the servers have a dedicated LAN for cross-communication for either SIP or SQL or should everything just go across DMZ lan?

6) Can I build my dialers to boot from LAN or is it basically a requirement to boot from local storage?

Thanks in advance for your help!

[geoff3dmg]

0) Master - Master MySQL replication will not work with ViciDIAL. Only Master - (Multi)Slave is supported.

1) Under moderate to high load the DAHDI dummy timer is not accurate enough. You need a hardware timer to solve this. If you are using ISDN then you can use the timers that come with that. If you are pure VoIP you'll need a dedicated timing card. I'm using Amfeltec cards.

2) My preference is Intel. The bottleneck tends to be the system bus anyway.

3) My preference is the distribute agents evenly between all telephony servers in the cluster (I use the phone aliasing feature). I'd also point out that out of the box (you can get round this with some dialplan work) agents can only manual dial and 3rd party transfer on a server with an active carrier setup. This may or may not be an issue depending on what you want to do.

4) It would depend on your concurrent call limits and the codecs you are using (5k calls with alaw/ulaw works out at about 800-900Mbit). Plus any other non-vicidial related traffic. I only use 10Gbe for my SAN interconnects.

5) See #4 regarding your SIP traffic. I don't know what SQL traffic is like. I've never had cause to measure it.

6) I assume you mean off a SAN? That's purely an OS issue, so I don't see why that wouldn't work.

[AkkerKid]

0) Sucks. Understood.
1) SKU-048-01 – x1 PCI Express System Timer?
2) Which is better? 8 Dialers with 4 cores each or 4 Dialers with 8 cores each? Performance vs Manageability?
3) Each box has direct connection the our SIP trucking provider. Is that what you mean?
4) Sounds like having bandwidth problems between servers is going to be the least of my concerns.
5) Cool.
6) I'll give it a shot.

Thanks Geoff3dmg!

[geoff3dmg]

1) Thats the one.
2) It's not really much more difficult to run 8 vs 4 diallers.
3) Yes.

[perci100]

How does a card like this compare to the USB voicetime stick? Also , would i need a timer in each telephone server? Previously i had thought these types of things were only for using pots/t1 .

[geoff3dmg]

According to matt the Sangoma USB timers can have issues with the USB bus drivers under linux (The timer just 'disappears'). Here's the forum thread about it (The whole thread is probably worth a read).

viewtopic.php?p=112588#p112588

[AkkerKid]

Two new questions:

A) Would a digium voice compression card per server would be a worth-wile investment? We are running a g729 shop with recording everywhere.

B) I have two clusters now. One seats 25 agents per dialer on single chip 4-core i-series CPUs and it's maxed out. The other seats about 35 agents per dialer with dual 6-core 3GHz Xeons and it's maxed out. Shouldn't I be able to put significantly more users on the machines with 12 cores in them? ( I'm assuming the the machines are getting maxed out because I'm starting to see MySQL connect errors and time sync errors now with no config changes...)

[mflorell]

Sangoma transcoding cards are higher density and more flexible than the Digium option. But you should take a look at each to figure out which would be better for you. Neither of them provide timing for Asterisk oddly enough.

[gardo]

We highly recommend using HAproxy (http://www.haproxy.org) for web load balancing. Very handy if you have more than 2 web servers. Set this up with Keepalived so you have a virtual IP address which handles all the requests and automatically load balance them to all your web servers.