Today, one of our network engineers, Chris Laffin, published a great post on the WordPress.com Developer Blog about how we use open source tools to analyze netflow data for our ever-growing global anycast network.
In June, I gave a talk at the dotScale conference in Paris about WordPress.com’s MySQL database architecture and infrastructure. The video is now online:
As you may have heard, on March 3rd and into the 4th, 2011, WordPress.com was targeted by a rather large Distributed Denial of Service Attack. I am part of the systems and infrastructure team at Automattic and it is our team’s responsibility to a) mitigate the attack, b) communicate status updates and details of the attack, and c) figure out how to better protect ourselves in the future. We are still working on the third part, but I wanted to share some details here.
One of our hosting partners, Peer1, provided us these InMon graphs to help illustrate the timeline. What we saw was not one single attack, but 6 separate attacks beginning at 2:10AM PST on March 3rd. All of these attacks were directed at a single site hosted on WordPress.com’s servers. The first graph shows the size of the attack in bits per second (bandwidth), and the second graph shows packets per second. The different colors represent source IP ranges.
The first 5 attacks caused minimal disruption to our infrastructure because they were smaller in size and shorter in duration. The largest attack began at 9:20AM PST and was mostly blocked by 10:20AM PST. The attacks were TCP floods directed at port 80 of our load balancers. These types of attacks try to fill the network links and overwhelm network routers, switches, and servers with “junk” packets which prevents legitimate requests from getting through.
The last TCP flood (the largest one on the graph) saturated the links of some of our providers and overwhelmed the core network routers in one of our data centers. In order to block the attack effectively, we had to work directly with our hosting partners and their Tier 1 bandwidth providers to filter the attacks upstream. This process took an hour or two.
Once the last attack was mitigated at around 10:20AM PST, we saw a lull in activity. On March 4th around 3AM PST, the attackers switched tactics. Rather than a TCP flood, they switched to a HTTP resource consumption attack. Enlisting a bot-net consisting of thousands of compromised PCs, they made many thousands of simultaneous HTTP requests in an attempt to overwhelm our servers. The source IPs were completely different than the previous attacks, but mostly still from China. Fortunately for us, the WordPress.com grid harnesses over 3,600 CPU cores in our web tier alone, so we were able to quickly mitigate this attack and identify the target.
We see denial of service attacks every day on WordPress.com and 99.9% of them have no user impact. This type of attack made it difficult to initially determine the target since the incoming DDoS traffic did not have any identifying information contained in the packets. WordPress.com hosts over 18 million sites, so finding the needle in the haystack is a challenge. This attack was large, in the 4-6Gbit range, but not the largest we have seen. For example, in 2008, we experienced a DDoS in the 8Gbit/sec range.
While it is true that some attacks are politically motivated, contrary to our initial suspicions, we have no reason to believe this one was. We are big proponents of free speech and aim to provide a platform that supports that freedom. We even have dedicated infrastructure for sites under active attack. Some of these attacks last for months, but this allows us to keep these sites online and not put our other users at risk.
We also don’t put all of our eggs in one basket. WordPress.com alone has 24 load balancers in 3 different data centers that serve production traffic. These load balancers are deployed across different network segments and different IP ranges. As a result, some sites were only affected for a couple minutes (when our provider’s core network infrastructure failed) throughout the duration of these attacks. We are working on ways to improve this segmentation even more.
If you have any questions, feel free to leave them in the comments and I will try to answer them.
We are looking at switching some of our servers from AMD Opteron Barcelona quad-core processors to the new Intel 5520 Nehalem processors. These are both 4 core CPUs, but the Intels utilize hyper-threading, so the OS sees 8 cores per CPU. It wasn’t that long ago that the first thing you did with a hyper-threading-enabled CPU was switch it off in the BIOS, but I have heard good things about Intel’s reincarnation of hyper-threading, so I decided to give it a shot.
I ran some real-world stress tests against these servers, adding them into the WordPress.com web pool and seeing how many requests per second they could serve before becoming 100% CPU bound effectively falling over. The types of requests served are varied; a lot are rendering web pages, but there are also quite a few image resizing operations thrown in here as well, as we spread this image work evenly over the 2500 cores in our web tier. Everything is php executed via fastcgi. I was a bit skeptical that there would be much of a difference between the two processors, but the numbers proved me wrong — the Nehalem’s are impressive.
2 x AMD Opteron 2356 Barcelona Quad-core 2.3Ghz
40 requests/second at 87.5% CPU utilization
2 x Intel 5520 Nehalem Quad-core 2.26Ghz
78 requests/second at 94% CPU utilization
Few things that I thought were interesting:
- On a per request basis, there isn’t much of a difference between the two. They both generate a given page in roughly the same amount of time.
- As CPU utilization approaches 100%, The Intel’s scale rather linearly, while the AMDs seem to struggle over the 85% range.
- The load averages were pretty high during these tests (35+ on the Intel box), but request times didn’t seem to suffer.
Has anyone else seen the same sort of results or maybe something to the contrary? These 2 configurations are roughly the same price, making it seem like a no-brainer to choose the Intels for web applications.
Since WordPress.com broke 10 million pageviews today, I thought it would be a good time to talk a little bit about keeping track of all the servers that run WordPress.com, Akismet, WordPress.org, Ping-o-matic, etc. Currently we have over 300 servers online in 5 different data centers across the country. Some of these are collocated, and others are with dedicated hosting providers, but the bottom line is that we need to keep track of them all as if they were our children! Currently we use Nagios for server health monitoring, Munin for graphing various server metrics, and a wiki to keep track of all the server hardware specs, IPs, vendor IDs, etc. All of these tools have suited us well up until now, but there have been some scaling issues.
- MediaWiki — Like Wikipedia, we have a MediaWiki page with a table that contains all of our server information, from hardware configuration to physical location, price, and IP information. Unfortunately, MediaWiki tables don’t seem to be very flexible and you cannot perform row or column-based operations. This makes simple things such as counting how many servers we have become somewhat time consuming. Also, when you get to 300 rows, editing the table becomes a very tedious task. It is very easy to make a mistake throwing the entire table out of whack. Even dividing the data into a few tables doesn’t make it much easier. In addition, there is no concept of unique records (nor do I really think there should be) so it is very easy to end up with 2 servers that have the same IP listed or the same hostname.
- Munin — Munin has become an invaluable tool for us when troubleshooting issues and planning future server expansion. Unfortunately, scaling munin hasn’t been the best experience. At about 100 hosts, we started running into disk IO problems caused by the various data collection, graphing and HTML output jobs munin runs. It seemed the solution was to switch to the JIT graphing model which only drew the graphs when you viewed them. Unfortunately, this only seemed to make the interface excruciatingly slow and didn’t help the IO problems we were having. At about 150 hosts we moved munin to a dedicated server with 15k RPM SCSI drives in a RAID 0 array in an attempt to give it some more breathing room. That worked for a while, but we then started running into problems where the process of polling all the hosts actually took longer than the monitoring interval. The result was that we were missing some data. Since then, we have resorted to removing some of the things we graph on each server in order to lighten the load. Every once in a while, we still run into problems where a server is a little slow to respond and it causes the polling to take longer than 5 minutes. Obviously, better hardware and reducing graphed items isn’t a scalable solution so something is going to have to change. We could put a munin monitoring server in each datacenter, but we currently sum and stack graphs across datacenters. I am not sure if/how that works when the data is on different servers. The other big problem I see with munin is that if one host’s graphs stop updating and that host was part of a totals graph, the totals graph will just stop working. This happened today — very frustrating.
- Nagios — I feel this has scaled the best of the 3. We have this running on a relatively light server and have no load or scheduling issues. I think it is time, however, to look at moving to Nagios’ distributed monitoring model. The main reason for this is that since we have multiple datacenters, each of which have their own private network, it is important for us to monitor each of these networks independently in addition to the public internet connectivity to each datacenter. The simplest way to do this is to put a nagios monitoring node in each data center which can then monitor all the servers in that facility and report the results back to the central monitoring server. Splitting up the workload should also allow us to scale to thousands of hosts without any problems.
Anyone have recommendations on how to better deal with these basic server monitoring needs? I have looked at Zabbix, Cacti, Ganglia, and some others in the past, but have never been super-impressed. Barring any major revelations in the next couple weeks, I think we are going to continue to scale out Nagios and Munin and replace the wiki page with a simple PHP/MySQL application that is flexible enough to integrate into our configuration management and deploy tools.