Tux the Linux Penguin

I work with Linux quite a bit at CG and it comes up in some interesting places. Consistently we find that the largest and smallest systems we work with are running Linux. If it’s a large system, like a database or storage network, Linux can provide the stability, robustness, and uptime required to make the project a success. If it’s a very small system, like an embedded display or a custom purpose device, we choose Linux because of its flexibility.

Over the last year or so, we’ve had a need to spin up many more virtual computers than we used to. It’s necessary to create a handful of  computers for each project we work on and the number of projects that we are taking on at a time is growing. Tools like EC2 and VMWare have made it very easy for us to create new virtual computers, but what about managing them? For Linux there are some great solutions. In this blog post I would like to talk about how we’re using some of the tools to deal with the large number of machines we are managing in Amazon EC2 for ourselves and our clients.

Cloud Control

The first part of the system is Amazon’s Elastic Compute Cloud (or EC2 for short). I’ve written about EC2 before, but to be really brief about it: EC2 allows you to create computers on the Internet quickly and lets you pay by the hour for their use. The instances you create are virtual machines that run in a large Xen based infrastructure that Amazon provides. To customize the software and configuration of the instances, Amazon lets you create snapshots of your computers (called Amazon Machine Images or AMIs) that you can launch new instances from.

It’s hard to say that making an AMI is difficult, it’s just a few keystrokes and a coffee break while the image is prepared and stored for your future use. When you begin to manage dozens and dozens of different images this becomes a problem. It became clear to us that we needed to have one AMI to manage, and build every computer from image dynamically. To do this we needed a way to pass specific information to each instance when it started up and then have a tool customize the instance based on the specific information about it. Amazon provides a method to pass specific information to an instance (it’s called user-data) and the good folks at Alestic have made some great AMIs around it with some good documentation.

Enter the Puppet Master

We have been using a tool called Puppet for nearly two years now to manage a handful of computers. We selected that to manage the whole lot of Linux computers that we were dealing with. Puppet allows us to describe how a computer should work in a general way. We can make collections of configurations for standard things so it’s easy to reuse what we create over and over again. Our Puppet configuration is stored in a Git repository so many administrators and developers can collaborate on it and our server configurations are backed up and version controlled automatically.

Bringing it all together

Puppet decides what configuration to use for a computer based on its hostname. We use the EC2 user data to pass a new instance the address of a Puppet server and the hostname that the machine should assume. When it boots up it sets its hostname and checks in to receive the configuration that we’ve stored for it. All changes to the machine happen through Puppet so we don’t have to spend a lot of time SSHing in and customizing the machine. It’s also very easy for us to duplicate a machine for testing or whatever we need.

While we originally did this to make our lives easier as we manage more machines there turned out to be some really cool side effects:

• Excellent Security: We don’t want to store sensitive information in Puppet. This means no passwords or secret stuff. To resolve this we require all developers and administrators to use key-based authentication to get access to the computers via SSH. This is very handy and it eliminates the need to remember passwords or for an administrator to have to reset passwords for users. Someday I would love to take the next step and have all of our users and machines be part of the Monkeysphere.
• Accountability: All of our configuration is tracked in a Git repository so we can see the history about what has changed on certain hosts and who changed it. Change control occurs automatically with Puppet and it’s easy to examine and understand what has changed and why it changed.
• Repeatability: By storing all Linux computer configuration in a single place we can easily repeat what we did for one server on another. Puppet institutionalizes all of our Linux knowledge in one place and saves us time every time we have to create a new machine. If you want to see how someone has done something in the past there’s no more need to dig through emails or documentation, just look at the facts in the Puppet repository and even copy and paste it into your new configuration.
• Portability: We use Puppet to manage much more than just EC2 instances. Physical machines and virtual machines in our ESX installation are managed this way too. It gives us one tool to take care of any Linux machine we deal with. Puppet also supports other operating systems. We’re looking to expand our use of it to Mac OS X machines and maybe even Windows sometime soon.

There are certainly other ways to solve the problems we’re up against, but this is the way we chose. We did some extensive evaluation of Chef (which is a tool like Puppet) and we’ve used Rightscale before. This sort of thing is becoming very important as we manage more and more computers. I expect we’ll see a lot of exciting products, techniques, and services in this space as time goes on.

If you have any questions or comments about our setup, or would like to discuss implementing something like this for your business, leave a comment or get in touch with us. We’d love to help.

Everyone seems to be touting HTML5 video as the “open” alternative to the proprietary Flash plugin required for .flv playback in the browser. But how open is H.264, the codec that powers HTML5 video, and the current pick for encoding video for online delivery? Using H.264 as the codec behind HTML5 video sours things a bit for me. H.264 is encumbered by software patents; to develop or distribute a player or encoder for H.264 you might have to pay a licensing fee to MPEG-LA. Even though MPEG LA announced last week (PDF) that H.264 will remain fee-less for free internet video through 2016, this is not the same as being free or open. MPEG-LA can still go after people that produce the software to encode or decode H.264. And MPEG-LA is not just one organization, it’s a collection of patent holders that have their own agendas.

All this is a bit of a slap in the face to the open standards that power the web. Imagine if you had to pay a half million dollars to create or display JPEGs, GIFs, or HTML… The only people that would be able to afford to make software for the web would be huge companies. But what are our alternatives? Beyond Ogg Theora and Matroska, the pickings are slim. These codecs are open and free, but not necessarily better than H.264. Plus it would be next to impossible to compete with the marketing machine of Apple behind H.264.

Open and free standards have been what has made the Internet successful since its inception. I think it’s important that users understand this so that the Internet of the future cannot be controlled by corporations with enough cash to cover licensing fees.

We have been doing some work with with Amazon’s Elastic Computing Cloud (EC2) which allows us to create virtual machines in the cloud in a few seconds. These are great for hosting websites, and what’s cool about them is that if you get Slashdotted or experience a similar unexpected spike in traffic you can create new hosts immediately. Recently Amazon added a new service called Elastic Load Balancing (ELB) which can distribute load across hosts. We’ve been looking at this for some of our recent development and infrastructure projects.

I just read this description of how ELB works by Shlomo Swidler from his Cloud Developer Tips blog. It’s a great reference.

You pay for ELB by usage just like everything else at AWS. From Amazon: “You are charged at $0.025 per hour for each Elastic Load Balancer, plus $0.008 per GB of data transferred through an Elastic Load Balancer.” For reference, on a deployment project in 2008 our Engineering team used a Cisco load balancer which I imagine cost a few thousand bucks.

Cost isn’t the only advantage. These can be created and destroyed quickly and remotely, allowing us to work more efficiently and spend less time visiting data centers in the middle of nowhere. This leads to improved quality of service for our clients as we can spend more time consulting on future technology growth plans and less time troubleshooting servers in cold, loud data centers.

This blog post brought to you by the iced coffee I am enjoying in the comfort and quiet of my office while deploying virtual machines!

This tool is called iozone. iozone is open source and runs on a ton of operating systems (including Windows). It runs several tests which can take some time to complete but provide the best overall view of the capabilities of a piece of storage. For instance, iozone runs a write test with files of different sizes and with different size records (the amount of data written at a time). It does this over and over again with writes, reads, random writes, random reads, and so forth. Since it’s running all these tests you can see what sorts of operations will have good performance and which ones will not perform so well. Check out the iozone documentation here.

One really great thing about iozone is that the output it generates can be easily placed in a spreadsheet program like Excel to generate a great 3d diagram describing your storage. Here’s a diagram I generated from some tests on a Linux server.

Results of a write test with iozone

This particular server performed quite well with large files and a record size around 1 MB (interesting to note, this is the same storage from the lmdd post. Notice that the parameters I tested with there are the same as the best write that this disk can do according to iozone!).

If you’ve been following my posts on storage performance testing I hope you’ve learned about some new tools that you can use to see what’s going on. I use these on every deployment to make sure we’re giving our clients solutions that they can depend for performance and reliability. As always, let me know if you have any questions about these tools. Happy testing!

bonnie++ is a series of file system tests that focuses on small files. It was designed to behave like a mail server does, creating and dealing with lots of small files (emails). bonnie++ is easy to run and outputs a CSV file that you can view with something like Excel. With the bon_csv2html command you can quickly generate html pages from the CSVs.

Here’s the output from bonnie++ running on a server:

The HTML output of bonnie++ on a Linux Server

At first glance the output can seem quite cryptic, but if we look close we can see that this provides us a great amount of information about latency and speed on different filesystem operations. I generally run this several times as I make changes to verify that the storage is providing the right performance characteristics. Tweaking a file system to make file system operations happen a few milliseconds faster may seem ridiculous, but in some environments it can make a huge difference.

Next time I’ll post about a tool that’s new to me but can test a disk in so many different ways I’m planning to run it on every system we install from now on.

Outside My Hotel in Malaysia – Do Not Feed The Monkeys!

I know the last time I posted here I said I’d be following up with another technical post, but instead I thought I’d share an experience I just had as I took a last minute trip for a client.

Normally if I take a trip it’s no big deal. I can write a blog post from where ever I go. My email is online, this blog is online, if I need to access something in my office I can just use our VPN to get connected. To use any of these I’d just need to have an Internet connection. Unfortunately, this wasn’t the case when last week I went to a fairly remote part of Malaysia.

A few coworkers and I were trying to make last minute adjustments to a product that one of our clients is launching. At first I wondered why even send us out there when we can get remote access or talk someone through it on the phone. When I arrived onsite I realized why this wasn’t an option — getting connected is near impossible there. We could head to a coffee shop and get some free WiFi, but with over twenty hops to servers in the United States and a twelve hour time difference, getting anything done was difficult.

The lack of connectivity was challenging. One of my responsibilities was interfacing with the local IT department and writing some scripts to integrate the client’s system with existing systems and processes. I quickly realized how much I depend on online references and documentation. When you can barely get connected to look up the answer to a question about syntax you really have to use your head. Not to mention, each software build for the project is about 300 megabytes and getting this from our office in New York was difficult and time consuming.

The idea of ubiquitous Internet connectivity is something that we take for granted. As connection speeds get faster and more reliable we lose efficiencies that we once had. I learned that the Internet is really an extension of my knowledge and a valuable tool that I need to do my job. Being cut off from it was an interesting and overall positive experience. Solving every problem by thinking and working it through was difficult and took more time, but genuinely figuring things out for myself was very rewarding.

Towards the end of my time there we found a cell phone store that sold GSM modems and prepaid 3G SIM cards that allowed us to get connectivity. While this does make the job a lot easier, I’m glad I had the experience of being mostly cut off from the rest of the net — something that will surely happen less often as the world becomes better connected.

I do some work on implementing storage systems for our clients, and we’ve found that different applications have different storage requirements. For example a video post production facility — like the facility at WWE — generally needs lots of disk space that is very good at reading and writing large files at high speeds. The storage here needs to provide good streaming throughput, because high quality video files generally have high bit rates, and are being stored or played back from the disk in real-time for ingesting, editing, or playout. If the storage system is not fast enough to read or write the file in real-time, frames will be dropped. This can cause unsatisfactory media files, programs to crash, or audio and video to become out of sync.

10,000 RPM SAS disks. That's some serious storage.

Suboptimal read/write performance can become a huge problem. When we put in a new system this is something we need to test. I usually do the test with a tool called lmdd.

lmdd comes from the lmbench tools which are provided by Bitmover for benchmarking systems. lmdd is great for testing streaming bandwidth. In most of our engagements with video, we install a Stornext or Xsan filesystem so we’ll run our tests against this. lmdd will probably work on any filesystem that you can mount on your Mac or Linux computer (Leave a comment if you need a version for Mac OS X, I have one compiled).  lmdd lets us verify exactly what the maximum number of megabytes per second we can push through the storage and point us to where we need to make changes to the hardware or software configuration. I use lmdd like this :

lmdd of=/path/to/test_file count=1g

lmdd if=/path/to/test_file

The first tests write performance and the second tests read performance. More information about the syntax is available in the manual page for lmdd. The results of the command from a server I was testing looked like this:

2147.4755 MB in 6.8003 secs, 315.7914 MB/sec

lmdd is great because it’s easy to read. This result shows I could write to the filesystem at 315 megabytes per second. That’s really fast! This is from a test with a server with a lot of RAM and a special filesystem that took advantage of that cache. When I run it on my Macbook, I get a result like this:

18342.6171 MB in 376.7685 secs, 48.6841 MB/sec

So the next time you’re interested in how your storage is performing give lmdd a shot and let me know how it goes. If you’re looking for more information about storage performance testing then stay tuned; I’ll be posting about testing storage with tools that benchmark small reads and writes next.

I was recently having a conversation with a friend and we both laughed when we thought back to the first five hundred megabyte hard drives that we had owned. Back then, the half-gigabyte drive was ridiculously expensive and physically huge. We both thought that it would be impossible to fill these drives up.

This of course was not the case. Now you’re lucky if an application can be installed in less than 500 MB, and as hard disk sizes grow, we find new ways to fill them up with applications, documents, and media. Digital files have become the most valued assets for most of our customers, so the organization, storage, and archiving of data is a serious concern.

I find that the best way to evaluate storage technologies solutions for our clients is to step back and take a look at the problems the client is looking to solve and the priorities dictated by their business needs. Usually, our clients’ storage needs require a combination of performance, reliability, disaster recovery, scalability, and manageability. Fortunately technology has stepped up to the challenge to handle the increased need for larger, faster, and more reliable storage.

Storage networking is a general term that encompasses many different technologies that provide excellent solutions to modern storage problems. A storage area network (SAN) is an architecture in which storage devices are connected in a high-speed, dedicated network and are presented to computers that are part of the same network. Using storage networking, we can accommodate our clients’ performance and reliability needs: by abstracting groups of hard drives as logical units (LUNs) we can stripe data across disks to increase speed and add redundancy by storing parity on the disks. This configuration will allow us to rebuild the LUN when a disk fails, without causing downtime or data loss.

Example SAN Configuration for Final Cut Pro Editing

A storage network abstracts the underlying hardware that provides storage services, providing some great advantages for disaster recovery. When we add tape libraries to a SAN we can make backups quickly and efficiently without slowing down the network or computers on it. We can also connect a SAN to another SAN that’s in a different building or even a different state. This allows us to easily replicate data to a secondary location so our clients can be up and running quickly if there is some kind of catastrophe in the data center.

Even the largest SANs will eventually get filled up with data. What happens when it’s time to increase capacity? With traditional storage, the system is shut down, new equipment is installed, and the data is migrated. This typically involves downtime and runs the risk of data loss if something goes wrong. With a SAN expansion is no problem. Since the storage services are abstracted from the storage hardware it’s easy to add capacity or replace older equipment, in many cases involving no downtime.

A SAN also provides centralized management for storage: administrators can look in one place to see the status of all storage in a data center.  This allows businesses to evaluate storage health and utilization, which can prevent problems and help plan for future growth.

As data becomes a more and more important part of business strategy, it becomes critical for businesses to have larger, faster, and more reliable storage services to keep things operating smoothly. Storage networking is a core component of these strategies. I’ll continue posting about our thoughts and experiences with SAN solutions, and try to shed some light on the storage ecosystem as new technologies emerge.