2008-06-25 09:56 PT

Today we are at Structure 08 and just saw Werner Vogels talk about Amazon WS. It was a great talk and really captured the problems that we set out to solve 4 years ag ao. An interesting statistic that he put forward is that, at scale, 70% of your engineering effort goes into infrastructure and 30% on application business logic. In our experience that percentage is on a sliding scale with the size of the user base. At start-up stage its less than 10%, at enterprise scale its 50%, at Amazon-Yahoo-Google scale its 70%.

The panel on moving to the cloud touched on a very hot topic: "Are we selling our soul with proprietary cloud providers". The poor Google guy was getting beat up the worst for the lock-in issue that Amazon, Salesforce and Google all have. A little unfair for roughing him up so badly on a common problem. Jason coined the term "Loving Cloud" there describing Joyent's hosting solution and the concept of a platform that is portable across providers. Our effort for the last 4 years has been around solving the problems Wergner pointed out but in transparent way that Jason is talking about. Overall a semi-heated discussion but it is great to see smart people debating these very important issues.

People really love the "cloud" buzzword. Its interesting that the underpinning computer science concepts, distributed computing, virtualization, autonomic computing aren't being mentioned specifically, but they are the real technologies that make this buzzword possible.

2008-06-23 12:26 PT

At the Velocity Conference now and actually hearing some cool things about performance and operations details. Here are some quick notes so far:

Jiffy performance measuring toolkit measure defined items, such as client side load for on-screen section not whole page.

jiffy.js - mark and measure concept

apache logging

ugly DB inserts.

firebug plugin

Operations is efficiently matching performance 
and reliability to users expectations.

20% of wiki pages slow down, lost 15% of overall traffic.  
 user drop off even on fast pages.

ads suck, load slower than page, not seen.
 - override document.write and document.writeln.
 - users click away before ad is displayed.

everyone follows common tier architecture.
 horizontal scale is model

sun is building benchmark tools

lower memory latency is key to processing performance.

old software doesn't scale on multi-core.
 Concurrency is issue.

1TB memory production scale system someday from sun.

Google noting that due to Moores law at some point soon,
 power will cost more than hardware with current trends.

Power provisioning playbook
 - consolidate workloads to fewest nodes needed for peak usage.
 - measure actual power consumption, specs innacurate.
 - trend watch to avoid oversubscription.

ISCA'07 report published by Google talks about this.

2008-05-22 17:01 CT

We're still here at the PHP|tek conference, having fun and connecting with people I haven't seen since last year. The theme for this years conference seems to be Rich Internet Application (RIA) technologies. To an extent its humorous to see the web starting to close the loop back to fat-client applications but it has the potential for great user experiences. I remeber when the specs for Adobe's Flex system came out. I was happy that someone was finally coming close to delivering on the promise of Java Applets. But it seems to have been a slower start for it than I thought. I really expected to see more fat client replacements in common usage by now using Flex. In any case, the conference is great and I've picked up a few jems. Morgan Tocker's MySQL talks I think were some of the best and, although I missed it, I hear Rich Bowen's new talk on Mod Rewrite was also really good.

2008-05-21 16:47 CT

Ryan, George and I are at PHP|tek this week. I'm giving a talk on Building Distributed Web Applications which is walk through of the major points of how we built our platform and tips and tricks to help with scaling and performance.

HostedLABS blog

General news, activities and thoughts on autonomic web oriented distributed computing.