|
One of our company’s early pioneers in web video, Light Reading, built its own video delivery system using two Windows Media Servers at a hosting facility. We asked Warren Hultquist to fill us in on the Light Reading system:
Our media servers have 73-GB drives partitioned for the OS and for data (about 57 GB per server after formatting). Because we encode at various bit rates (for users with different connection speeds), storage is a challenge. We also keep log files for media stats tracking. Since storage is cheap, our plan was generally to just get bigger drives and/or add NAS.
For bandwidth, we have 100-Mbps connections for each of the two live servers. The file sizes are 15 to 50 MB—depending on the bit rate and length of the video (we use 256-Kbps and 768-Kbps files, running anywhere from two to eight minutes in length).
The servers mirror each other; we simply upload video to one and the file propagates to the other. One server is in a hosting facility in New Jersey, the other is in San Jose, Calif., so we get geographic failover.
The hosting company technicians keep the servers, load balancer, and firewall up and running, and they run the backup processes. We perform OS patches. We monitor the storage capacity and maintain the tracking software (NetTracker).
We don’t load balance. In the beginning we did a round-robin, with every fourth request going to the second server. Once we saw that the main box could handle the load, we turned off the round-robin and now we just use it for straight-up failover.
We’re serving roughly 350 video clips. Last year, we served 110,321 total sessions, 141,353 total downloads, 887,449 MB transferred, 302 average sessions per day, 2,116 average sessions per week, 9,193 average sessions per month, 387 average clips downloaded per day.
This doesn’t even break a sweat for our server. Our bottleneck would be concurrent connections rather than processing power. With a 100-Mbps connection, we guessed we’d have three connections per megabyte with network overhead, so 300 concurrent connections max. We’ve probably seen about 50 concurrent connections max.
|