GARDOL

Adaptive Denial of Service Attack Mitigation

by John Walker

 

Questions to Answer

When an apparent distributed denial of service attack hits your site, it's easy to panic, especially if it effectively knocks your site off the Web and the boss is running around like a beheaded chicken urging you to "do something". This is even more severe at sites like Fourmilab, where I have to play both the rôle of the chicken and the beleaguered system administrator simultaneously. Still, before you can ``do anything'' that's likely to improve the situation, you need to carefully determine the facts of the matter, which process you can begin by answering the following questions.

Are we really under attack?

You may have decided your site is under attack because you've observed a large bulge in the hit rate on your server, seen your inbound and/or outbound network bandwidth hit the peg, or watched the CPU load on your server reach saturation.
Web access log: January 2004
An attack on Fourmilab.
The first question to ask yourself is whether these symptoms are actually the result of a DDoS attack or something else entirely. Certainly, all of these things are characteristic of a DDoS attack, but they do not, by themselves, indicate an attack is actually underway. The chart at the right shows the onset of the DDoS attack against Fourmilab in January 2004. Prior to the attack, the hits per day had been at the typical level of 500,000 to 650,000 per day. As the attack began, hits per day increased to over a million per day and stayed at that level. Analysis of the request log showed the additional hits were completely stylised requests for nothing but the home page, originating from thousands of different IP addresses all over the world, with each requestor repeating at a more or less constant rate, and several hundred new IP addresses "recruited" into the attacking population each hour. The attack requests were easily distinguished from legitimate requests for the home page because they contained request header variables which no browser was known to send. In this case, an unexpected bulge in request rate did indeed indicate an attack. But this is not always the case.

Consider the access graph at the left, covering a week in April 2003. Once again, requests were running at the typical rate with the usual slowdown during the week-end when, on the 13th of April, the hit rate exploded to more than double the usual average rate.
Web access log: April 2003
Not an attack.
This hit the outbound bandwidth hard--its 2 Mb/sec capacity, which is about twice the average bandwidth requirement, saturated, and response time went from the typical 100-500 milliseconds to in excess of five seconds for many requests. The hits just kept on coming at this rate for the first hours of the 14th as well, at which time they finally began to taper off. An attack? Nope. Examination of the request log immediately revealed what was going on. One of the documents on our site had been mentioned on one of those sites frequented by fat people who cannot spell, who click like Skinner-pigeons on any link posted there, in order to download, but of course never read, the contents before posting their erudite "arguements" against its (unread) content and disdain for the author of its (unread) words. You can usually identify the source of this kind of traffic bulge from the "Referer" field in your HTTP log. (If you don't use the "combined" log file format which includes this field along with the "User-Agent" which identifies the browser, here is an excellent reason for starting to do so.) In this case, most of the initial hits to the document in question has a Referer pointing to a posting on a site for iconoclasts who all think alike, and a visit there confirmed that as the source of the hits. One could then breathe a sigh of relief, "this shall soon pass", and indeed, once the posting scrolled off the home page of the site, the hit rate returned to normal. A few minutes of research put an end to the panic over a potential attack and pointed to patience as the proper prescription.

One month earlier, in March 2003, after a typical week and week-end, the request rate started climbing from its typical rate to higher and higher levels on each successive day, reaching
Web access log: March 2003
An attack, but not on Fourmilab.
more than a million hits a day by the 21st. These hits originated from all over the world, and were primarily requests for images from Earth and Moon Viewer which, requiring substantial CPU time on the server, vastly exceeded the server's CPU capacity, sending its load average, which typically peaks around 3.5 to 4, to more than 400 processes waiting to run. The referer field in the first hit of these requests were mostly blank or from search engines. Was Fourmilab under attack? No--Iraq was! The outbreak of hostilities in Iraq causes tens of thousands of people all over the world, having seen satellite images of Iraq on television, to run to their computers to look for a source, whereupon Earth and Moon Viewer popped out at the top of the search. I suspect most of these visitors didn't realise that the images they were requesting were generated from a static imagery database, so no matter how far they zoomed in, they wouldn't be able see the tanks roll toward Baghdad.

What is the kind of attack?

DDoS Attacks of the First Kind: Nuisance Level

DDoS Attacks of the Second Kind: Outbound Bandwidth / Server Saturation

DDoS Attacks of the Third Kind: Inbound Bandwidth Saturation

Where is it coming from?

What is the profile of the attacking hosts?

Is the attacking population static or dynamic?

What is the trend of the attack?

What are the attackers sending?

Is our inbound bandwidth being saturated?

Is our outbound bandwidth being saturated?

Is our server being overloaded?

Are the attack requests stereotyped?

Can attack requests be distinguished from legitimate ones?

What is the request trajectory of an attacker?

When did the attack start?

Who were the first identifiable attackers?

Download Gardol:   gardol.tar.gz source archive

Read Gardol source code


Fourmilab Home Page

by John Walker
March 2004
This document is in the public domain.