PHOBOS - Re: drbd (was nmbd)

Ben Lindstrom mouring at pconline.com
Wed Nov 10 00:08:10 MST 1999



On Tue, 9 Nov 1999, banibrata dutta wrote:

> hi,
> 
> 99% in agreement with Ben.
> 
> Every client that we have to our HA-Solaris box (that runs dedicated Core 
> Telecom s/w), asks about the wire-cut problem, the partitioned- cluster 
> problem. Well, if you ask me, can't the 4 port NIC go kaput, well, then let 
> me ask you, can't the motherboard and all the gents sitting on it, memory, 
> processor, bus-wiring go kaput...?? the thing
Or your massive $10k fibre switch.. Or in a case at my job recently a
CISCO router motherboard.  

> is PROBABLITY!! No HA solution is 100% HA. In a duplex-hot-standby HA 
> solution what happens in case of a double-fault... at some point HA
> ends...
> 
I was only ensuring that one does not depend on the other 3 network
channels as their only "Fail over" route.  You still need to double
up the card or use a serial line.  

Besides Linux still may go insane if a hardware vanishes from under it, or
the fact some motherboards will freak if a card burns out.

> Also, I have one more thing to say. Although many good folks on this list 
> are actively doing very good work in HA, but I think, the work is little bit 
> biased-towards/focussed towards web-servers (and DNS-servers) only... I am 
> designing a VoIP Signalling gateway that needs to be HA... and to minimise 
> call-losses, and call-disconnection and failures in feature provisioning 
> throughput, availability of redundant link, and state synchronization (at 
> app level) is a must. In this case, if the things are bad at NIC level yet 
> link is OK, the other box detects the failure, resolved partitioned cluster 
> (to some extent, and in 90% of the cases), and takes-over. So, the level of 
> trust, in NIC and machine hardware is much more than connectors, and 
> cables... Doesn't anyone in this list have such requirements as I have...? I 
> just wanted to know, who are working towards applying Linux-HA in areas 
> other than web-servers, and mail-servers... The mission-criticality I feel 
> is much more in my case... (of course this is my PoV).
> 
Basicly your working on the same problem an ex-employer was working on.
Supporting 4 T1s (later 8 T1s) over 4 primary Interactive UNIX boxes and
two m88k Data General boxes with shared intelligent SCSI arrays.   Looking
back they were pretty primative failovers, and we easily could screw it
up. (Their main biz was Call-Back, IVR, and were forging into VoIP when 
they went bell up.)

> There is a point of disagreement with Ben though, i.e. the extent to wihch i 
> can, and are willing to move the stuff/logic into hardware. Hardware has 
> COSTS... so it's a critical balance (to be figured out by the final HA 
> implementor/designer/planner), between the software components and the 
> hardware components he wants to plug-in. So the design of HA needs to be 
> such that modules/solutions can be plugged in, i.e. if i want i may add 
> SSLized card, or use it as a software module, if i want link-aggrregation on 
> card - i go out buy one and use it, else i put in a software based 
> link-aggregation (ethernet link level load-sharing, balancing, with higher 
> aggregate bandwidth). Adding a PHOBOS, SSL card, and 4 port FT NIC might 
> cost quite high, and make the Linux-HA solution "not-so-interesting" a 
> thing.
> 
As long as the 4 port card is not the single method of doing HA over I
have no problems.  It would be interesting to bond all 4 ports in 100bt
full duplex mode to see what speed I would get out of it.  But if I want
to improve speed I would rather look into Fibre cards then channel bonding
which is less stable and more head aches when it comes to debugging.

SSL cards are interesting if your doing SSL websites (or daring and do
SSL VPN) and if you have the money and require the extra power are good
things, but I would perfer to see hard facts showing that the SSL card is
worth the cost of it.

But I suspect that most webserver clusters (or single machines) if speced
out right don't run at 100% load for over 10% of the time.  If they do 
then either one needs to profile their code or upgrade the box.  (Yes
fokes your old 486s are cool, but you don't run 100,000 hit a day
webservers off them without seeing a performance hit. =)

Basicly it boils down to.. If the 4 port card does something that you
need.  Great.. But as long as you don't build a HA around it and come
crying to me when it blows and you lose billions due to it.=)  I would
rather stick to dual 100baseT Full or Fibre cards.  




More information about the Linux-HA mailing list