July 2010 Apache Traffic Server benchmark

I reran my benchmarks with the latest "trunk" of Apache Traffic Server, to make sure we're not regressing. I also tweaked the number of worker threads a little, a gut feeling tells me that with Hyper Threading, our auto-scaling algorithm isn't optimal (and, it really isn't). Here are the latest numbers, running over a GigE network (two Linksys el-cheapo switches between clients and server)

3,160,237 fetches on 3,666 conns, 1,800 max parallell, 1.58012e+09 bytes in 30 seconds
500 mean bytes/fetch
105,341.10 fetches/sec, 5.26704e+07 bytes/sec
msecs/connect: 1.46781 mean, 6.674 max, 0.093666 min
msecs/first-response: 16.3333 mean, 615.34033 max, 0.121333 min


That is, 105k QPS (with keep-alive) for small objects, over the network. It's pushing 52MB of payload at this speed, but remember the average size is very small (500 bytes). My box is an Intel i7 920, Quad core.


Why Traffic Server defaults to not allow forward proxying

We have discussed numerous times on the Apache mailing lists about the reasons why Apache Traffic Server ships with a default configuration that is almost entirely locked down. Our argument has been that we want to assure that someone testing TS is not accidentally setup in the wild as an open proxy.

I recently moved all of www.ogre.com to be served via a TS reverse proxy setup. Within minutes from setting it up, and while watching the log files, I found entries like these in the logs:

1273927263.211 0 ERR_CONNECT_FAIL/404 485 GET http://proxyjudge1.proxyfire.net/fastenv - NONE/- tex -
1274057848.081 0 ERR_CONNECT_FAIL/404 485 GET http://proxyjudge1.proxyfire.net/fastenv - NONE/- tex -
1274236765.403 4 ERR_CONNECT_FAIL/404 485 GET http://proxyjudge1.proxyfire.net/fastenv - NONE/- tex -

Of course, my server doesn't allow this, since it's setup to only accelerate www.ogre.com traffic. But case in point is, I think we've done the right thing of shipping Apache Traffic Server with a very restrictive configuration.


Ogre is now running on Apache Traffic Server

I've just switched over to serve all of www.ogre.com out of Apache Traffic Server. The site is still managed and created using Apache HTTPD, PHP and Drupal, but that is running as an "origin" server to ATS. This gives me a few benefits over serving straight out of Apache HTTPD:

  • Static content is automatically "cached" on the ATS server, and it can serve such content very fast with low latency.
  • I can jack up keep-alive much higher than I dared doing with HTTPD. Fwiw, I still use the pre-fork MPM, so I have limited number of processes and can't afford to tie those up with idle KA connections.
  • In a pinch, I could turn the HTML generated from Drupal to be cacheable, and serve straight out of ATS. I'm contemplating making this setting automatic, so when the load on the box hits a certain level, all HTML will also be cached by ATS. That would increase my capacity by at least a magnitude I think.

This change required no changes on my Drupal site, but I did change the port on my Apache HTTPD virtual host:


    ServerName www.ogre.com

I then installed Apache Traffic Server to listen on port 80, and I also told it to only bind a specific IP on my server (I have three IPs for different things). I also increased the RAM cache size and Keep-Alive timeouts, so I now have these changes in etc/trafficserver/records.config:

CONFIG proxy.config.proxy_name STRING kramer3.ogre.com
CONFIG proxy.config.http.server_port INT 80
CONFIG proxy.config.http.keep_alive_no_activity_timeout_in INT 60
CONFIG proxy.config.http.keep_alive_no_activity_timeout_out INT 1
CONFIG proxy.config.http.transaction_no_activity_timeout_in INT 15
CONFIG proxy.config.http.transaction_no_activity_timeout_out INT 30
CONFIG proxy.config.cache.ram_cache.size LLONG 33554432

LOCAL proxy.local.incoming_ip_to_bind STRING

Next, I added a disk cache to use for ATS, etc/trafficserver/storage.config:

/disk/tmp 134217728

This creates a 128MB cache in /disk/tmp. I know, very small, but this is still experimental. Finally, I added a remapping rule to etc/trafficserver/remap.config:

map http://www.ogre.com/ http://www.ogre.com:8080/

After starting everything up, the entire site is now reverse proxied (or accelerated) through Apache Traffic Server! As you can see, the changes necessary to ATS are fairly small, and pretty straight forward, most of the default settings 'just work'. It's a miracle.


Adobe Photoshop CS5 vs The Gimp

I've only used CS5 for a few hours, but I have to say, this beats The Gimp hands down in every way... The content aware fill is by far the coolest thing I've seen in a long time, and the ease of making a lot of my "common" tasks (for example Unsharpen mask on luminosity only) is awesome. Don't get me wrong, The Gimp is great, particularly with the Scheme scriptability, but there's no question Photoshop / CS5 is vastly superior. I'm still running the trial version on my Windows7 VirtualBox installation, and so far, it rocks.

Update: Yes, there is a similar plugin for the Gimp...


First "alpha" release of Apache Traffic Server

We've finally produced our first "alpha" release of Apache Traffic Server. It can be fetched from your local Apache Download Mirror. This first version, v2.0.0-alpha, should be reasonably stable as it does not contain a ton of improvements over the old Yahoo code. The 2.0.0 releases also only supports Linux, but a number of 32-bit and 64-bit distros have been tested.

We're hoping to get some testing done on this release in the next week or two, so we can make the "final" 2.0.0 release. After that, the plan is to aggressively start making "developer" releases of trunk, which has some impressive improvements (like, up to 2x the performance in some cases). Apache Traffic Server is adopting the same versioning scheme as Apache HTTPD, so, v2.0.x is a "stable" release, while v2.1.x is a developer release. This implies, of course, that the next stable release will be v2.2.0, and we're hoping to get that out the door sometime this summer!

People interested in Apache Traffic Server are highly encouraged to join our mailing lists (see the incubator page), or come join us on #traffic-server on freenode.net.


Synergy is dead? Long live Synergy+!

I used to use Synergy on my Linux desktop and Mac laptop, to allow me to share the same keyboard and mouse between the two. After upgrading to a 64-bit Linux distro (FC12 btw), Synergy would no longer compile. Instead of trying to fix this, I went searching, and found Synergy+ instead. It works great on my boxes, and is even backwards compatible with Synergy. But best of all, it's actively maintained and supported, give it a try.

Now that I found it, I realized that it's even available in the Fedora Core distribution (yum install synergy-plus)


Traffic Server performance

I just got a new shiny desktop at home, a Quad Core i7 920, running 64-bit Linux 2.6.32 (stock Fedora Core 12). Nothing fancy, it's a $250 CPU, but I did some artificial tests of Traffic Server on it just to see where we are on a modern machine and kernel. The tests are done between two Linux boxes, both on GigE network, and with two Linksys switches in between the two boxes. So, this is definitely not a production quality network in any way. I ran with keep-alive enabled, doing 100 requests per connection, each request fetches a 500 byte body out of cache (i.e. 100% cache hit ratio). I know, this is a completely unrealistic test, but it is still interesting to see what we can do. Here's the best run out of three:

2270442 fetches on 23464 conns, 2000 max parallell, 1.13522e+09 bytes in 30 seconds
500 mean bytes/fetch
75681.40 fetches/sec, 3.78406e+07 bytes/sec
msecs/connect: 0.3878585 mean, 11.1295 max, 0.089 min
msecs/first-response: 16.3695 mean, 459.379 max, 0.1295 min

So, over 75,000 requests per second, at 16ms latency, and I think some of that latency can be attributed to the two switches (wish I had a better setup). An interesting side note, the Traffic Server process can actually use about 470% CPU on this Quad Core box, so an extra 70% of a CPU is "gained" here from Hyper Threading. How does this compare to my old desktop? Well, the old system was a Core2, and in the same test, it was able to pull off around 35k QPS.

Anyways, these results aren't too shabby, and we've just started!

Caveat: This is using the "trunk" version of Traffic Server, the first release that is soon coming out won't be quite this fast.

Update: I updated the numbers with the results after forcing my CPU to run at highest CPU frequency at all time. No overclocking though, this is a standard i7 920 setup.



Modifying a Linux software RAID mirror

While upgrading my desktop to Fedora Core 12, I also decided to modify a software RAID mirror that I had on the old system. In particular, I wanted to do two things, without losing data:

  1. Change the partition layout, the old disks had partitions for OS and swap allocated to them, which I no longer needed. I know, LVM would have been nice here, but alas, I wasn't using it on this older box.
  2. I also wanted to upgrade to EXT4, and I've heard that upgrade process can potentially corrupt the entire disk (ask Bryan Call if you don't believe me).

So, since I'm using mirroring, I figure the "right" approach would be to just break the mirror, and do a migration safely that way. And of course, that does work, and there's plenty of information in the man-pages how to do this, but I figure I'll write down the steps I took so I can remember it myself (and maybe someone else finds it useful too).

The first step is to break the mirror, in my case, the RAID mirror is /dev/md2 using /dev/sdb and /dev/sdd, and then mount the broken mirror half on the file system:

# mdadm --manage /dev/md2 --fail /dev/sdb1
# mdadm --manage /dev/md2 --remove /dev/sdb1
# mkdir /mnt/old-data
# mount /dev/sdb1 /mnt/old-data

We now have a complete "copy" of the mirrored data, make sure /mnt/old-data (or whatever you named it) looks good. The next step is to get rid of the old RAID device, and create a new one. In my case, I need to run fdisk as well, to change the partition layout, but I'm not going into details here exactly what I did, since it's very specific to my setup.

# mdadm --stop /dev/md2
# fdisk /dev/sdd
# mdadm --create /dev/md0 --level 1 --raid-devices=2 missing /dev/sdd1

Finally, we can now create the new filesystem, and migrate the old data over to it:

# mke2fs -t ext4 -j /dev/md0
# mkdir /mnt/new-data
# mount /dev/md0 /mnt/new-data
# rsync -av /mnt/old-data/ /mnt/new-data

The final step is to add back the "old-data" partition into the new mirror, but before doing so, make sure the "new-data" looks alright.

# umount /mnt/old-data
# fdisk /dev/sdb
# mdadm --add /dev/md0 /dev/sdb1


That's it! Easy, and very little risk for lost data or corruptions.


Fedora Core 12 with nVidia drivers

I fried my old CPU while benchmarking Traffic Server (yeah, it's that fast ...), so I spent most of the day replacing the mother board / CPU. I figured I'd take the opportunity to ugprade to Fedora Core 12 as well. This mostly worked well, with one exception: FC12 now comes with the OpenSource nvidia drivers, and they are slow. Unfortunately, these drivers also conflicts with the nVidia closed source drivers (which I install from source), causing the installation to fail. But, with a few changes in my system configs, I worked around it.

First, change /etc/grub.conf and edit the "kernel" line, to include something like


Secondly, edit /etc/modprobe.d/blacklist.conf and add a line

blacklist nouveau

I'm not sure why both are necessary, maybe I did something wrong and you only need one or the other. Let me know if that is the case, I'm happy with what I got, and not wasting any more time on it.



Subscribe to Ogre.com RSS