Hardware Load Balancers + High Availability

Forums › IT Pro and developers › Hardware Load Balancers + High Availability

tigercorp

621 posts

Ultimate Geek
+1 received by user: 54

Subscriber

Topic # 86488 7-Jul-2011 10:55

Hi.

Can someone please give me a quick run down on how to avoid a single point of failure when using a hardware load balancer?

Is it simply a matter of having 2 of them?

Quick background - we've always load balanced our web app using MS Network Load Balancing which also provides a measure of HA.
While our parent company has always recommended a hardware balancer as well, none of our customers have every chosen that route. Now someone wants to do it so and I need to get my head around what the architecture should be, with HA taking priority over performance.

Ragnor

8027 posts

Uber Geek
+1 received by user: 387

Trusted

Subscriber

Reply # 490810 7-Jul-2011 21:36

Need more details, before you get to load balancing for web servers do you have?

- 2 Internet connections
- 2 routers with some kind of failover or balancing enabled
- 2 primary network switches

If not, you should look at hosting in commercial data centre as those are equally as important points of failure as a web server.

Regarding MS NLB it's good for the price (ie: nothing it comes with windows) but has a pretty dumb balancing method, for example it will keep sending traffic to a server that is up even if IIS is frozen or unresponsive or experiencing degraded performance. Hardware balancers have a lot more bells and whistles eg: smarter balancing methods, caching, compression, session options etc.

ptinson

677 posts

Ultimate Geek
+1 received by user: 27

Trusted

Reply # 490828 7-Jul-2011 22:12

What Ragnor said. However the question was more specific to the actual devices.

It depends on what your customer is really asking for and what level of HA you are after, however strictly as far as hardware load balancers go most are active-standby, you can run active-active but it tends to bite you in the backside if you aren't careful.

I know of at least one vendor that is playing with hot standby (my term as i cant remember what they actually said) where by you have a number of devices that are active for different functions and a shared fail over device.
You have to work out what contention ratio you are happy with but the idea has merit in reducing standby space heater sprawl.

If HA is all you really want and you don't need the performance i would also consider the software appliances, just be careful as some are more usable than others.

But like I say, it depends on what your customer wants and why.

Paul

meat popsicle

tigercorp

621 posts

Ultimate Geek
+1 received by user: 54

Subscriber

Reply # 490856 7-Jul-2011 23:36

Thanks for the replies, guys.

The requirement still stands though - how to remove the single point of failure.
I know I don't know enough to know what question to ask, so I'll try to explain with the following example.

If I've got 3 web servers sitting behind a hardware balancing appliance and the appliance dies, its all over.
To remove this single point of failure should I:

- have another complete 3 web servers with another hardware balancer in front of them, with the two load balances sharing a virtual IP. So essentially 2 complete systems in active-passive.

- have just another hardware balancer that can make use if the existing web servers, again with the two load balances sharing a virtual IP. So in this case the appliances are only active-passive.

My guess is that its the second case, but I'd like confirmation or alternatives.

More details - the app is a medical one and must be hosted onsite (where the ever increasing data/images are stored) and it will be a small cluster of servers sitting behind a hardware balancer. A small number of users + the balancer and cluster take care of the performance aspect. Its the HA I'm worried about.

insane

2281 posts

Uber Geek
+1 received by user: 370

Trusted

Subscriber

Reply # 490869 8-Jul-2011 00:29

There are different ways you can do this, however to cover both the load balancing and provide extra HA we (Maxnet) provide the following setup:

We have two types of Load balancers (F5 BIG-IP series), the one type (local traffic managers) are a pair working in an active/standby (HA) and balance traffic between a pool of servers using a virtual IP address per service being balanced. When a server in the pool fails its given health check it is removed from the pool and traffic is only balanced across the remaining servers which are still available.

Now on top of that we also have another set of load balancers which balance traffic between multiple locations / pools, (Global traffic managers), so if for example all the servers in the load balanced pool are not available, the traffic will be directed at a different virtual IP address at another location.

eg If all a customers load balanced servers in Auckland do down, then we throw traffic to Christchurch, assuming they have a primary + DR service.

It's the same system which Trademe use for their load balancing

ptinson

677 posts

Ultimate Geek
+1 received by user: 27

Trusted

Reply # 490906 8-Jul-2011 08:35

Option 2 if its all on site and you don't need site fail over and the application isn't a lifeline service.

If its a life line service then option one with some additional bits.

You mentioned large images, if the application is web based you can use an F5 to compress these, it works very well.
Does the application have any quirky session persistence requirement, this can also dictate what hardware vendor you choose.

I have used most hardware load balancers, from Cisco to F5, and i would have to say that for shear flexibility the F5 wins hands down, you can really do some powerful tinkering right up to layer 7.

Of course the choice of vendor comes down to $$ as well.

Paul

meat popsicle

Ragnor

8027 posts

Uber Geek
+1 received by user: 387

Trusted

Subscriber

Reply # 491079 8-Jul-2011 13:52

The next inevitable question is what's the budget / how much can you spend?

tigercorp

621 posts

Ultimate Geek
+1 received by user: 54

Subscriber

Reply # 491158 8-Jul-2011 16:00

I know site failover will be out of budget so I'll be presenting my option 2 (insane's option 1) as the preferred solution with MS NLB as the cheap alternative.

The actual hardware selecton will be made by the customer, but I'm guessing (ie I haven't done any real reading yet) that most of these types of appliances wil have application/IIS aware smarts?

Ragnor

8027 posts

Uber Geek
+1 received by user: 387

Trusted

Subscriber

Reply # 491162 8-Jul-2011 16:08

Another option between hardware load balancers and MS NLB in price that might suit your situation is buy two extra entry level servers (eg: Dell RS210's or something like that around $1000-1500ish ea) and install HAProxy on them.

You can read about HAProxy here
http://haproxy.1wt.eu/

There is a list of companies who use HAProxy here (eg: Twitter, Redit, Stackoverflow, Github etc)
http://haproxy.1wt.eu/they-use-it.html

So you'd have two cheap dedicated "hardware" servers in active/standby doing the load balancing running a proven "software" load balancer solution. in front of your 3 web servers.

HAProxy has a lot more features that MS NLB, doesn't cost the earth like F5 and Cisco appliances. Might be a good compromise option to suit the budget.

insane

2281 posts

Uber Geek
+1 received by user: 370

Trusted

Subscriber

Reply # 491306 8-Jul-2011 23:36

tigercorp: I know site failover will be out of budget so I'll be presenting my option 2 (insane's option 1) as the preferred solution with MS NLB as the cheap alternative.

The actual hardware selecton will be made by the customer, but I'm guessing (ie I haven't done any real reading yet) that most of these types of appliances wil have application/IIS aware smarts?

Out of interest where are your web servers located? If they are hosted in a DC then check with them to see whether they offer load balancing as a service, might save you some CAPEX that way. If you're dead set on purchasing tin and are tight on budget then you could look at Foundry Server Iron SLBs, they do what most people need from an SLB.

You also say that a fail over site will be out of the question... you could quite easily get a cheap VM hosted at another site or even another provider to use as a fall back / DR. You don't have to have all your servers with the same provider to make use of Global/Site load balancing services.

You could even just get a $5 hosting somewhere and put up a splash screen for your users so they don't get greeted by a 404 if your primary site goes down.

tigercorp

621 posts

Ultimate Geek
+1 received by user: 54

Subscriber

Reply # 492158 11-Jul-2011 18:51

insane:...

Out of interest where are your web servers located? If they are hosted in a DC then check with them to see whether they offer load balancing as a service, might save you some CAPEX that way. If you're dead set on purchasing tin and are tight on budget then you could look at Foundry Server Iron SLBs, they do what most people need from an SLB.

You also say that a fail over site will be out of the question... you could quite easily get a cheap VM hosted at another site or even another provider to use as a fall back / DR. You don't have to have all your servers with the same provider to make use of Global/Site load balancing services.

You could even just get a $5 hosting somewhere and put up a splash screen for your users so they don't get greeted by a 404 if your primary site goes down.

The servers have to be located on the customer's site, where the data is as its a web based app as opposed to a (fairly static) website.

Thanks for the ideas and help everyone!

insane

2281 posts

Uber Geek
+1 received by user: 370

Trusted

Subscriber

Reply # 492242 11-Jul-2011 22:46

Right I see,

Does the customer have dual WAN connectivity with auto fail-over? If not then perhaps a hosted solution could be seen as a good value proposition and offer the redundancy they clearly would like.

Having said that, load balancing servers hosted in a DC is not too different to balancing two servers on the end of a data circuit/s.

Hit me up with a PM if you think you may want to outsource it.