I was reading about security stuff in the .NET framework, and dealing with cryptographic classes in it, and it sort of set me wondering. Here are all these different encryption classes, with different block and keys sizes, cipher modes, all that jazz – but what are their performances like? Specifically, I’d read something saying how some ‘weaker’ encryption algorithms are better (in some speed-critical applications) ‘cos they’re faster. I wondered how much?
Thus, I decided to benchmark the Symmetric alogrithms in the .NET Framework – DES, Triple DES, RC2 and Rijndael. To make life interesting, I thought I’d try them with differenct key sizes and block sizes, and cipher modes.
So, I’ve linked to definitions of these factors, but for those who don’t want to read vast chunks of Wikipedia, here are my (simplified) definitions. For anyone really interested in learning how to program with encryption properly (and in learning why their 128 bit key probably isn’t 128 bits strong) I can strongly recommend the book ‘Practical Cryptography’ by Bruce Schneier and Niels Ferguson.
Symmetric ciphers are ones like you used when you were a kid. You have some operation that turns a message into garbage, and then the reverse of that operation turns that garbage into a message. Some algorithms don’t have a reverse – they are asymmetric ciphers, and are a whole different kettle of fish.
Keys are the password you use with your cipher. For example, if you’re cipher as a kid was to shift all letters in the alphabet, then the key might be the numbers of characters shifted. Big keys are harder to break. Think of it as being just like a password or PIN number. If I tell you that my PIN is 4 digits, you might be tempted to guess all 10,000 possibilities, and on average you’d figure my PIN out after 5000 tries. If my PIN was 8 digits, then there is 100,000,000 options – and you’re less likely to try all the possibilities, eh?
Block sizes. Well, okay, some ciphers work on blocks of data, rather than each byte (or each ‘letter’). These are block ciphers. There are also stream ciphers, where each byte is encrypted one by one. Anyway, in block ciphers there is a limit to how much data can be encrypted without ‘leaking’ information. Larger block sizes can encrypt more data without that leakage. (That’s not to say that the block has been decrypted, but an attacker could start to learn things about the contents of that block.)
Cipher modes don’t really have a parallel with how you did codes as a kid. I guess I would describe it that if the cipher is about how you make an apparently random set of bits, then the cipher mode is about how you then use them. There are lots of different modes, but the .NET framework classes only seem to support 3 – ECB (Electronic Cookbook), CBC (Cipher block chaining) and CFB (Cipher Feedback).
So, what are the algorithms:
- DES – An old encryption standard, now regarded as offering poor security, but so widely used that it is still in operation as a legacy system.
- Triple DES – An improved version of DES, made by essentially applying the DES 3 times.
- RC2 – A moderately old encryption algorithm. Flexible key lengths, but short block size.
- Rijndael (aka AES) – The latest encryption standard. The Rijndael algorithm was selected from several as part of a competition. It wasn’t regarded as the most secure, but it was quite quick. The Advanced Encryption Standard (AES) is actually a subset of Rijndael.
I found a nice text file – “The complete works of Shakespeare” – as my test data.
For each algorithm, for each mode, key and block size, the test program encrypted and decrypted the data twenty times, and reported the average ‘time’ for each operation. I was using the Win32 QueryPerformanceCounter function, which doesn’t really return a time so much as cycles. However, all the tests were done on the same machine, so they’ll do just fine for comparison purposes.
With the several factors tested, there are many ways of slicing the data. It’s worth noting that these results are pretty rough, as the times taken also include file IO operations, and with any modern PC there’s also something else happening at any single time. Also, the times are the total time taken to encrypt and decrypt, which might not be the same for each operation. Treat the results as a loose guide.
First let’s look at the raw results. You can get the results here (Excel file).
Unsurprisingly, DES is fastest – given it’s age, and the low level of security it offers now. Triple DES with the longest key it supports was generally slowest. RC2 covered the full range of results, which is also unsurprising, given it’s flexibility, and Rijndael sort of falls in the middle.
The first thing I noticed was how few tests there were using DES or Triple DES. RC2 and Rijndael are much more flexible in their use.
Next, it’s interesting to note that RC2, DES and Triple DES using Cipher Feedback Mode (CFB) were all very, very slow. They all seem to suffer very badly using CFB.
So, excluding the CFB results then (as they are so exceptionally slow), what do the other results show? Well, Rijndael does not suffer so badly in CFB, although CFB is slower.
ECB appears slightly slightly faster in the table, though examining the CBC Mode Graph shows little difference.
To compare the modes, I looked at just the operations done with the Rijndael cipher.
Again, we see little difference between ECB and CBC, so I guess there’s no reason not to use the more secure CBC mode over ECB (for an example of it’s weakness see here). Also, for 128 bit blocks (as required by the AES standard), CFB is as quick as ECB.
Rijndael is not great in CFB mode with blocks of longer than 128 bits.
Okay, so let’s focus on just one mode (CBC) and look at the results shown in the CBC Mode Graph. Well, it’s interesting to note that RC2 with a 112 bit key was quite quick – faster than with some shorter keys. However, it’s only about 6.5% longer to use 128 bit Rijndael – which is a key that is 14% longer. Doubling the key with Rijndael to 256 was only 10% longer than 128.
Longer blocks take longer to encrypt and decrypt. 64 bit blocks seems a little short these days, only being safe for up to a couple of hundred megabytes. 128 bits seems more reasonable. 256 bits seems excessive. Rijndael seems to have little penalty for using 256 bits over 128, though if you do, you’re not using an AES standard encryption.
DES and Triple DES are old. DES isn’t secure, and Triple DES doesn’t seem to offer much given Rijndael and RC2 being much faster than it.
In terms of cipher modes, these classes only seem to support ECB, CFB anc CBC. ECB is generally regarded as being a poor mode – it’s not very secure. CFB was typically slower than CBC, and as Microsoft have already implemented the classes, some of the advantages CFB (i.e. encryption and decryption being identical operations) have been lost.
So, then examining Rijndael in CBC mode, well, there is little penalty for using 256 bit keys or 256 bit blocks. However, it’s probably worth sticking to 128 bit blocks as 1) it is plenty, and 2) it is AES compatible.
All in all, I was surprised by how similar a lot of the results were for different algorithms, and I was surprised by how slow some of the CFB mode operations were.
To be honest, I can’t really think of a reason not to use Rijndael with 128 bit blocks, in CBC mode. Unless time is a really critical factor, 256 bit keys are stronger. Finally, the RijndaelManaged class in the framework is a managed class, rather than a wrapper for a COM object.
So, the winner is Rijndael!
So, yes, when I get around to it.