This year's Internet Problem Solving Contest is over. This year was special for me, because for the first time I was the root responsible for making sure that shit doesn't happen during the contest and in case shit does happen, making it go away.
Historically, we always had server outages during the contest due to the relatively high number of HTTP requests from contestants, especially at the immediate beginning of the contest, when everyone is downloading the problem statements, and then also as the end draws near, when everyone keeps spamming F5 on the results page.
After I took charge of our servers a year and a half ago, I decided to put an end to these outages. Unfortunately, last year I was busy around the time of the contest (writing my Bachelor's thesis), so I didn't really have the time to tune our servers. This year, however, I finally got around to boosting the performance of the IPSC website.
Tuning the server
Usually the server outages were caused by Apache spawning too many server processes and causing the server to swap itself into oblivion. The previous version of the web interface was a PHP thing and it used to be hosted on our main server that we use for everyday tasks, SVN/git hosting, people's personal pages, SSH access etc.
Due to the fact that the server hosts people's personal pages, often using PHP and things like .htaccess for mod_rewrite and other stuff, we can't really ditch Apache on that server, and getting that mammoth of a server to behave is no easy task. (It's no secret I have little love for Apache, or even PHP, for that matter.)
Last year, Tomi Belan rewrote the whole web frontend in Python as a WSGI application. Personally, I consider this a step in the right direction, obviously, with my involvement in the Python web development community. My immediate intention was to put this app on a separate virtual machine behind nginx and tune it to accept more than just a few (<10) requests per second.
So that's what we did. We used a spare virtual machine that had been sitting on our physical server unused for many years, and set up nginx and uWSGI to run and serve the app. Tomi implemented certain measures to make the most expensive pages load much faster (the results pages).
Then, about a week before the actual contest, I did some benchmarks to find out how high a load the server will be able to handle. I played around with httperf, autobench and ab for a while, and I found out that if the server was hit with too many requests (about 500 per second), it would just respond with a 502 error to most of them. I found that weird, I certainly didn't expect the server to swat nearly all requests for a page that takes about one millisecond to render away with a 502 Bad Gateway error.
So I tried to look at the logs, and in the nginx error log I found the following:
connect() to unix:///run/uwsgi/app/ipsc/socket failed (11: Resource temporarily unavailable) while connecting to upstream
The first candidate for a solution I found by googling was to increase the net.netfilter.nf_conntrack_max sysctl. However, that didn't help, so I kept searching until I found the real cause.
Each socket keeps a queue of incoming connections that have not yet been accepted by the application listening on the socket. By default on Linux, the queue is limited to 128 incoming connections, which, I think, is surprisingly low. uWSGI sets it even lower for its socket, it only allows 100 incoming connections. In uWSGI the relevant setting is listen, which I raised to 2048. However, this alone isn't enough as Linux won't allow any process to raise this above the global limit, which is set in the net.core.somaxconn sysctl whose default value is, as I already mentioned, 128. So I raised this to 2048 as well.
Next, nginx started to complain about insufficient worker connections:
2013/06/01 00:23:50 [alert] 24814#0: 768 worker_connections are not enough 2013/06/01 00:23:50 [alert] 24814#0: *14708 768 worker_connections are not enough while connecting to upstream, client: 220.127.116.11, server: , request: "GET / HTTP/1.1", upstream: "uwsgi://unix:///run/uwsgi/app/ipsc/socket:", host: "ipsc.ksp.sk"
This one was fairly straightforward to diagnose, I just had to raise worker_connections and that was about it.
The story is not over yet, though. After fixing this, the server started to spam the error log with these:
2013/06/01 00:51:52 [alert] 29982#0: *42773 socket() failed (24: Too many open files) while connecting to upstream, client: 18.104.22.168, server: , request: "GET / HTTP/1.1", upstream: "uwsgi://unix:///run/uwsgi/app/ipsc/socket:", host: "ipsc.ksp.sk" 2013/06/01 00:51:52 [crit] 29982#0: accept4() failed (24: Too many open files)
This is a symptom of an insufficient limit on the number of open file descriptors per process. If you try to google for these error messages, you'll come across a lot of suggestions to modify /etc/security/limits.conf, raise the system-wide limit on open files to 4096 or whatever (both soft and hard), add pam_limits.so to the common-session config file of PAM, logout/login and restart the server. While this may work for you the moment you start the server manually, the higher limit won't apply if the server is started at bootup by init, since in that case, it doesn't go through any PAM mechanism which would apply the settings from limits.conf.
Instead, the actual solution on Debian is to enable ULIMIT="-n 4096" in /etc/default/nginx or add the following to your nginx.conf:
After going through all this setup, I was satisfied with the server setup. It could serve about 600 requests for the main page per second, and buffer a sufficient number of requests before responding with 5xx codes instead of just failing immediately when the number of requests increased.
I also did a few benchmarks on the main results page. The page itself renders within 10 milliseconds, which is quite fast, considering it's half a megabyte long. That means, this page responds at a rate of more than 100 requests per second, which should be sufficient, as this corresponds to a bandwidth of about 62 MB/s – that's about one half of the bandwidth of the server's gigabit connection.
The morning before the contest
The day of the contest I got up at about 9:00 a.m. The contest itself was starting at 12:00 p.m. in our timezone. I ate my breakfast, checked my email, read the RSS and was getting ready to go out to take care of a few things before the start of the contest. As I was about to pack my laptop, I noticed that the SSH connection to our server timed out. This is nothing exceptional – our faculty, where our servers are housed, has connectivity outages quite often.
Unfortunately, as I was leaving, Usamec told me the IPSC server was down as well, while other machines in the same room were accessible. Crap.
So I hurried to school to threaten the server with a hammer to make it behave. As I was on my way, Miňo Plžík called me and said he was investigating the cause. When I arrived to school and connected a display and keyboard to the machine, I was greeted by this backtrace:
Jun 8 09:16:22 prezident kernel: [4545085.998861] vif vif-10-1: vif10.1: Frag is bigger than frame. Jun 8 09:16:22 prezident kernel: [4545085.998941] vif vif-10-1: vif10.1: fatal error; disabling device Jun 8 09:16:22 prezident kernel: [4545085.999041] BUG: unable to handle kernel NULL pointer dereference at 00000000000008b8 Jun 8 09:16:22 prezident kernel: [4545085.999157] IP: [<ffffffff8134d981>] _raw_spin_lock_irqsave+0xe/0x25 Jun 8 09:16:22 prezident kernel: [4545085.999254] PGD 466ab067 PUD 4d7ee067 PMD 0 Jun 8 09:16:22 prezident kernel: [4545085.999334] Oops: 0002 [#1] SMP Jun 8 09:16:22 prezident kernel: [4545085.999390] CPU 2 Jun 8 09:16:22 prezident kernel: [4545085.999413] Modules linked in: xt_tcpudp xt_physdev iptable_filter ip_tables x_tables xen_netback xen_blkback xen_gntdev xen_evtchn xenfs bridge stp lm85 hwmon_vid ipmi_devintf ipmi_msghandler loop snd_pcm snd_page_alloc snd_timer rade on coretemp iTCO_wdt acpi_cpufreq mperf crc32c_intel snd ghash_clmulni_intel aesni_intel aes_x86_64 aes_generic cryptd ttm drm_kms_helper drm power_supply i2c_algo_bit i2c_i801 soundcore iTCO_vendor_support processor button i2c_core evdev pcspkr thermal_sys ext4 crc16 jbd2 mbcache dm_mod raid1 usbhid hid md_mod sg sr_mod sd_mod cdrom crc_t10dif ahci libahci ehci_hcd usbcore e1000e libata usb_common scsi_mod [last unloaded: scsi_wait_scan] Jun 8 09:16:22 prezident kernel: [4545086.000551] Jun 8 09:16:22 prezident kernel: [4545086.000573] Pid: 3342, comm: netback/2 Not tainted 3.2.0-4-amd64 #1 Debian 3.2.39-2 FUJITSU PRIMERGY TX100 S3/D3009-A1 Jun 8 09:16:22 prezident kernel: [4545086.000738] RIP: e030:[<ffffffff8134d981>] [<ffffffff8134d981>] _raw_spin_lock_irqsave+0xe/0x25 Jun 8 09:16:22 prezident kernel: [4545086.000859] RSP: e02b:ffff8801d5e9bc30 EFLAGS: 00010002 Jun 8 09:16:22 prezident kernel: [4545086.000936] RAX: ffff8801a9990218 RBX: 00000000000008b8 RCX: 0000000002d4792a Jun 8 09:16:22 prezident kernel: [4545086.001039] RDX: 0000000000010000 RSI: ffffc900124d2000 RDI: 00000000000008b8 Jun 8 09:16:22 prezident kernel: [4545086.001142] RBP: 00000000000008b8 R08: 0000000000020055 R09: 000000000000003b Jun 8 09:16:22 prezident kernel: [4545086.001244] R10: ffffffff81433340 R11: ffffffff81433340 R12: ffff8801a999b740 Jun 8 09:16:22 prezident kernel: [4545086.001347] R13: 0000000000000000 R14: ffff8801a999b818 R15: ffff880028e60002 Jun 8 09:16:22 prezident kernel: [4545086.001453] FS: 00007f7aa20d5700(0000) GS:ffff8801ef300000(0000) knlGS:0000000000000000 Jun 8 09:16:22 prezident kernel: [4545086.001565] CS: e033 DS: 0000 ES: 0000 CR0: 000000008005003b Jun 8 09:16:22 prezident kernel: [4545086.001650] CR2: 00000000000008b8 CR3: 000000003f10f000 CR4: 0000000000002660 Jun 8 09:16:22 prezident kernel: [4545086.001753] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 Jun 8 09:16:22 prezident kernel: [4545086.001856] DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400 Jun 8 09:16:22 prezident kernel: [4545086.001959] Process netback/2 (pid: 3342, threadinfo ffff8801d5e9a000, task ffff8801dfba47f0) Jun 8 09:16:22 prezident kernel: [4545086.002080] Stack: Jun 8 09:16:22 prezident kernel: [4545086.002111] 0000000000000000 ffffffffa03d4ddd ffffffff81433340 ffff8801a999b740 Jun 8 09:16:22 prezident kernel: [4545086.002240] 0000000002d4792a 0000000002d4792a 0000000000000000 0000000002d4792a Jun 8 09:16:22 prezident kernel: [4545086.002363] ffff880028e60002 ffffffffa03d4f71 ffff8801a999b740 ffffc90011837580 Jun 8 09:16:22 prezident kernel: [4545086.002486] Call Trace: Jun 8 09:16:22 prezident kernel: [4545086.002522] [<ffffffffa03d4ddd>] ? xen_netbk_schedule_xenvif+0x35/0xd6 [xen_netback] Jun 8 09:16:22 prezident kernel: [4545086.002631] [<ffffffffa03d4f71>] ? netbk_tx_err+0x3f/0x4b [xen_netback] Jun 8 09:16:22 prezident kernel: [4545086.002727] [<ffffffffa03d5517>] ? xen_netbk_tx_build_gops+0x59a/0x9e2 [xen_netback] Jun 8 09:16:22 prezident kernel: [4545086.002840] [<ffffffff81004be5>] ? phys_to_machine+0x13/0x1c Jun 8 09:16:22 prezident kernel: [4545086.002924] [<ffffffff810040b9>] ? xen_mc_flush+0x124/0x153 Jun 8 09:16:22 prezident kernel: [4545086.003007] [<ffffffff81003124>] ? xen_end_context_switch+0xe/0x1c Jun 8 09:16:22 prezident kernel: [4545086.003094] [<ffffffff81003b77>] ? xen_mc_issue.constprop.24+0x31/0x49 Jun 8 09:16:22 prezident kernel: [4545086.003191] [<ffffffff8100d02f>] ? load_TLS+0x7/0xa Jun 8 09:16:22 prezident kernel: [4545086.003265] [<ffffffff8100d6a6>] ? __switch_to+0x13b/0x258 Jun 8 09:16:22 prezident kernel: [4545086.003343] [<ffffffff810359b7>] ? arch_local_irq_enable+0x7/0x8 Jun 8 09:16:22 prezident kernel: [4545086.003430] [<ffffffff81039834>] ? finish_task_switch+0x88/0xb9 Jun 8 09:16:22 prezident kernel: [4545086.003517] [<ffffffff8134c60c>] ? __schedule+0x5ac/0x5c3 Jun 8 09:16:22 prezident kernel: [4545086.003596] [<ffffffffa03d5a8d>] ? xen_netbk_kthread+0x12e/0x6f5 [xen_netback] Jun 8 09:16:22 prezident kernel: [4545086.003702] [<ffffffff8105f943>] ? add_wait_queue+0x3c/0x3c Jun 8 09:16:22 prezident kernel: [4545086.003785] [<ffffffffa03d595f>] ? xen_netbk_tx_build_gops+0x9e2/0x9e2 [xen_netback] Jun 8 09:16:22 prezident kernel: [4545086.003896] [<ffffffff8105f2f1>] ? kthread+0x76/0x7e Jun 8 09:16:22 prezident kernel: [4545086.003972] [<ffffffff81354ab4>] ? kernel_thread_helper+0x4/0x10 Jun 8 09:16:22 prezident kernel: [4545086.004057] [<ffffffff81352b73>] ? int_ret_from_sys_call+0x7/0x1b Jun 8 09:16:22 prezident kernel: [4545086.004146] [<ffffffff8134dcbc>] ? retint_restore_args+0x5/0x6 Jun 8 09:16:22 prezident kernel: [4545086.004232] [<ffffffff81354ab0>] ? gs_change+0x13/0x13 Jun 8 09:16:22 prezident kernel: [4545086.004307] Code: be 01 00 00 00 e9 cf fe ff ff e8 bb 32 d2 ff 85 c0 0f 95 c0 0f b6 c0 c3 e9 92 32 d2 ff 53 48 89 fb e8 58 32 d2 ff ba 00 00 01 00 <f0> 0f c1 13 89 d1 c1 e9 10 66 39 ca 74 07 f3 90 66 8b 13 eb f4 Jun 8 09:16:22 prezident kernel: [4545086.004859] RIP [<ffffffff8134d981>] _raw_spin_lock_irqsave+0xe/0x25 Jun 8 09:16:22 prezident kernel: [4545086.004952] RSP <ffff8801d5e9bc30> Jun 8 09:16:22 prezident kernel: [4545086.004999] CR2: 00000000000008b8 Jun 8 09:16:22 prezident kernel: [4545086.030275] ---[ end trace ad4724e43471a71d ]---
While I was on my way, Miňo found out that the problem was most likely caused by this bug in the “Xen made easy!” bug tracker. Later I found the corresponding Debian bugreport. In short, it's a bug in the Linux netback/netfront implementation used by Xen, which is triggered when the DomU's kernel tries to offload the fragmentation of a packet bigger than 64 KB to the virtual network interface. When this happens, the netfront driver running in Dom0 disables the affected network interface altogether with no way to restore it, other than shutting down the DomU and creating it again. In addition, it occasionally tries to dereference a null pointer, which has other side effects on the whole Xen stack.
Anyway, according to the first bug report, a workaround should be to disable GSO offloading in the affected DomU, so we rebooted the whole physical machine (because of the kernel oops), started up all DomUs and disabled GSO.
This was the >40 minutes long outage you might have seen during the practice session.
The contest itself
Now for the thing we were the most curious about: how the server would handle the load during the actual contest.
I made a plot of the number of requests per minute throughout the contest.
The beginning at 12:00 p.m. was the first big challenge. Just to be on the safe side, though, Mišof cooked up two password-protected zipfiles containing all the problem statements that people could download from another server and at 12:00 he published the password. This presumably reduced the load on the main server a little bit.
Still, at the beginning of the contest, the number of requests spiked at more than 5200 per minute. The server didn't really notice, though, with a load of about 0.25. To put this into perspective, on the server hosting the IPSC website previously, at this point the load would jump into the double-digits and it would be swapping like crazy. If there were any significant lags this year, then they can only have been caused by insufficient bandwidth of the internet connectivity.
Within a few minutes the request rate dropped significantly and settled slightly above 1000 requests per minute where it would have stayed throughout the whole contest, had things gone without any unexpected trouble. However, as anyone who's been around computers knows, shit just does happen.
At 12:13:33, the requests just stopped coming in. That was really weird, I was still connected to the server from the local network, the public network interface appeared to be up, only no packets were going through. At first I was somewhat clueless, however, the syslog on the hypervisor revealed the issue to be the same one that hit us in the morning:
Jun 8 12:13:33 prezident kernel: [ 8053.515765] vif vif-4-1: vif4.1: Frag is bigger than frame. Jun 8 12:13:33 prezident kernel: [ 8053.515838] vif vif-4-1: vif4.1: fatal error; disabling device
What the hell? How? I thought I had implemented the workaround (disabled GSO) on all relevant DomUs. As it turns out, nope, I forgot about this particular virtual machine. As a matter of fact, it didn't even have ethtool installed. A quick shutdown, xm create, try to install ethtool. However, as soon as the machine booted up, it was flooded with incoming HTTP requests from contestants nervously spamming the F5 button, which meant the interface crashed before I could successfully both install ethtool and disable GSO. (I guess I'm too slow at typing. d-: ) That's the small spike you can see in the plot. So, to get the thing to run again, I had to give it another reboot.
Finally, I was able to get the machine to run with disabled GSO. I thought we were safe and that there was nothing worse that could happen to us. Again, I was proven wrong when the exact same thing happened at 12:30:53.
This time, I was really desperate. We had already used the workaround suggested in the “Xen made easy!” bug report and still it didn't help. I was considering the following two options: either try to migrate the whole virtual machine to a physical computer as fast as possible, or try to perform a dist-upgrade – the machine was still running Debian squeeze, I thought that Wheezy might already contain the fix mentioned in the bug report. (This was before I found the Debian bug, which makes it clear this wasn't true.)
After a short while, Miňo suggested trying to disable every possible kind of offloading on the network interface with the following short shell script:
Being as desperate as I was, I set this as the post-up script of both ethernet interfaces and rebooted the machine once again. From then on, I was constantly watching the HTTP logs for any silent period longer than a second. We even had a real-time visualization with Logstalgia on a projector. Luckily, after that incident, the interface hadn't crashed again.
The last big challenge was the end of the contest when everyone is reloading the results page. This is somewhat different from the beginning of the contest, because in the beginning, everyone wants to download the problem statements in the first possible second, which means a large number of requests in one second, but the content is mostly static. At the end of the contest, though, the load tends to build up more gradually as more and more teams give up during the last few minutes and just keep reloading the results page. Also, the results page is generated dynamically.
Actually, to be honest, the actual contents of the results page are only generated every 15 seconds, not on each request, but still, this page has to go through a little bit of Python processing before it is returned to the client.
Anyway, again, there was a spike at about 3500 requests in the first minute after the contest which faded pretty quickly and, again, the server didn't even notice – its load was lower than at some points during the contest.
To sum up, I think the contest went fairly well, at least from my point of view. There was the hiccup caused by a kernel bug and partly by myself, which I apologize for. However, for the rest of the contest, the server handled the load above any expectations.
Looking forward to IPSC 2014 already. (-;
As a bonus, feel free to watch this Logstalgia visualization of the contest.