Null Disquisition

A friend and I have been bouncing around the idea of a caching system that ran on Amazon’s cloud for a while now. Basically something like memcached, but without the (very real) limitations of physical memory or the need of a whole server. Sure, it’s hard to beat the speed of memory-level read access, but I think the appeal of a distributed, limitless cache might outweigh the slowdown.

Idea

Provide an interface for storing/retrieving serialized data on S3

Pretty simple idea, pretty simple implementation. Thanks to the S3 interface provided by Boto, things were a lot easier. I’m going to keep this open source under the MIT license. You can check out the code on GitHub repository – please feel free to fork, improve, submit, etc.

Overview

A quick walkthrough of the code will reveal truly how simple this is. The Client class provides basic CRUD methods for interfacing with S3: put, get, update, delete. The put and update methods store a timestamp as the “expires” header for the file to keep track of cache expiration. Also these two methods write a “type” header to the meta-data so CloudCached knows how to de-serialize the file.

class Client:
"Here's the class schema"
    def get(self, key)
    def put(self, key, value, time_to_expire=3600, replace=False)
    def update(self, key, value, time_to_expire=3600)
    def delete(self, key)

There are 6 basic data types used in this code for serializing any bit of python data: basestring (for str and unicode), int (for int and long), complex, float, and other. The other data type represents anything that is not a base type in Python. These “other” types get pickled while everything else just gets str’d.

The put method checks the md5sum to make sure everything went through cleanly (maybe a bit costly, but worth it in my opinion). cPickle is used in favor of pickle for obvious reasons (it’s much faster).

Results

Some very early tests show that this might just be usable.

    CloudCached Benchmarks (10 runs)
    --------------------------------------------------------
    Test                 |  Average (s)     | Total (s)

    --------------------------------------------------------
    GET integer          |  0.0283360004425 | 0.283360004425
    GET string (32 byte) |  0.0315794944763 | 0.315794944763
    GET string (512KB)   |  0.1265994787220 | 1.265994787220
    PUT integer          |  0.0650457143784 | 0.650457143784
    PUT string (32 byte) |  0.0563205003738 | 0.563205003738
    PUT string (512KB)   |  0.1773290872570 | 1.773290872570
    --------------------------------------------------------

Advantages

• Highly distributed. S3 data is distributed across multiple availability zones and could therefor be utilized by an application running across multiple availability zones.
• No size limit. Unlike the physical limitations of a memcached machine (or cluster of memcached machines), S3 does not have limits on the number of files (caches) you can store. Also, with S3, you can write files from 1 byte to 5 GB (although I think a 5GB cache file would defeat the purpose).
• Parallel read access. If applicable to the application, cache reads can be largely parallelized which could potentially give linear speedup to the cache loading.
• No server necessary. Since the application is reading and writing directly to S3, there is no need to a “cache server”. This could lead to a great deal of savings for people running multiple memcached machines. Memcached servers typically have a large memory capacity which means a m1.xlarge or c1.xlarge EC2 instance (assuming it’s running in EC2).

Considerations

It’s going to be hard to beat the speed of memcached. As far as speed is concerned, I’m using built-in Python stuff including urllib, httplib, xml.sax, etc (all of which are used by Boto). It might be worthwhile to write a C implementation of the S3 communication methods (but maybe not). The most costly part of this code aside from network communication is probably the serialization, and since cPickle is used there is not really improvement to be made there.

It might be cool to couple the meta-data with SimpleDB.

I registered cloudcached.com in case this gains some momentum. I will post updates and benchmarks there as they arrive.

-David

After a few days of tinkering with EC2MPI, I spent some time polishing up a stat mech MPI simulation. The code in question is a 2d Ising model simulation using Replica Exchange. Right now it stands at around 400 lines of C++ using STL vectors (which I love). Once I know it works (or at least works well enough) I might post it up here, but for now I’m just trying to generate pretty hysteresis plots and observe the critical behavior of a 2d Ising model system. Here’s a picture with points on it.

I leave the interpretation to you. The best part of this is that I can do these MPI runs without burning a hole in my lap (the MacBook gets rather warm). -David

Addendum to two previoius posts.

The other day, I noticed my 8GB USB volume that I use for temporary incremental backups was quite full. Curious, since the folders I back up to that volume do not total but 200MB or so, and rsync was supposed to be doing incremental backups (link-dest ftw).

After a little searching around, I found someone who had a similar problem (and a solution). When you format a volume with OS X it will, by default, ignore file ownership (the linked article explores why this is perhaps). This proves to be a problem for rsync which considers file permissions and ownership as part of the file stat (as it should). Luckily the fix is easy – “Get Info” for the volume in question, then at the bottom unselect “Ignore ownership on this volume”

You will probably want to delete any backups that have been created (since they won’t have the correct file ownership). Source: Terminalapp.net