Go big, or go home!

I’m currently working on setting up Wowza on an EC2 “Cluster Compute Quadruple Extra Large” instance (or as I’ve heard it called, the “super-duper-quadruple”, which sounds like something I’d get at Five Guys). There’s no pre-built AMI for this one, so you have to use a stock Linux image (I use the standard Amazon one) and install Wowza with a subscription license, and do the tuning yourself. But the payoff is this: for $1.30 an hour, you get a streaming server capable of delivering 10Gbps of data.  On a 750Kbps stream, that’s over 13,000 concurrent clients. This for about the same cost as nine or ten m1.small instances which can deliver an aggregate of about 1.5Gbps. On a reserved instance, you can get this down to just under 75 cents an hour.

In addition to Ludicrous Speed on the network I/O, this instance comes with 8 multithreaded Xeon 5570 cores (at 2.97GHz), 23 GB of RAM, and 1.7TB of local storage. (a quick speed test downloaded a half-gigabyte file in about four seconds, limited by the gigabit interface at the remote server). This is roughly equivalent to a moderately configured Dell R710. There’s also a GPU-enabled version of this that adds a pair of nVidia Tesla GPU cores.

If that’s not enough, you can go bigger, with 16 cores, 60GB of memory, and 3.5TB, Recently, someone clustered just over a thousand of these instances into the 42nd largest supercomputer in the world.

As of right now, these monster instances are only available in the us-east-1 zone.

 

 

9 Comments On “Go big, or go home!”

  1. Isnt 10gbitps just something you can reach between AWS instances in the same availability zone? Can you really confirm that this speed is achievable with actual live streaming clients?
    I see that you have done some wokr on this, so can you tell me, what is the maximums speed you have been able to reach with small and large instances?

    Reply

    • Well, the paradox is that the only way you can speed/load test an instance that big is from another server with similar capabilites – those are few and far between, if you get outside Amazon’s cloud. I have heard from other Wowza clients and the engineering staff at Wowza that it is in fact possible to get the full 10G out of it from all the streaming clients (although there’s a limitation in Java that kicks in around 5G, but I understand it’s relatively easy to work around)

      On small instances (m1.small), the practical speed limit is about 150M, m1.large can do about 250M.

      Reply

      • Thanks for info. The small instance seems to be a beter choice than large taking into ccunt price/performance ratio (price difference is 4x).

        Reply

        • Generally speaking, yes, multiple m1.small (150Mbps) is a better option than m1.large (250Mbps) or m1.xlarge (350Mbps), if you’re using the devpay license. However, on the daily subscription, you’re paying per instance, so a cc.4xlarge (10Gbps) is a much better option, as it’s about 9-10x the cost of an m1.small for 60x the bandwidth. 

          Reply

    • on windows cc1.4xlarge i have only 1Gbs nics

      i simulate 1200 connections to live with 2 instances same zone and get 1Gbs traffic!

      I try to figure what’s wrong!

      Reply

  2. This looks like a great solution for an event I have coming up.  Ian, have you actually used this instance for streaming?   also did you have the Transcode module on ? 

    Reply

  3. Ian how did you tune the javaVm for that large instance?

    Reply

  4. Hi,

    i ran some stress test on 2 cc1.4xlarge spot instances. one wowza server and 1 flazr to make connections on my live stream…

    running windows 2008, the nic is 1gbs i was expecting 10gbs!
    and test show 1gbs bw!….

    do you get 10gbs ?

    Thanks

    Daniel

    Reply

Leave a Reply

Your email address will not be published.