Showing results for 
Search instead for 
Do you mean 

Subversion 1.7: Zippy, but should you upgrade?

by Dreamer on ‎11-25-2011 01:34 PM

by Lutz Dornbusch, Senior Consultant, Polarion Professional Services Racecar with Subversion 1.7 paint jobSubversion 1.7 is a release mostly devoid of new features, but with many performance improvements. From some users’ highly appreciated WC-NG to mostly invisible HTTPv2, the question is what to expect from these improvements. Because Polarion’s application lifecycle management products are built on top of Subversion, making SVN a core issue for our customers, we felt the need to highlight and benchmark some of the most visible performance boosts of SVN 1.7. I conducted tests with several key SVN operations and found some results impressive enough to suggest that upgrading to SVN is probably worthwhile for most users. But there are some important caveats as well, especially on the client side (at the time of this writing). In this article, I'll share the results of my performance evaluation of Subversion 1.7 and recommendations for upgrading. If you like, you can download a PDF of this article (open download, no registration). Before you start reading - some handy links I'd like to share with you:

Measurement Procedure

I measured mostly on a small 1.6 GHz dual-core laptop. The reason is that even if I have faster hardware for production, I want to see performance under low-tech conditions. But I also included some datasets on a high-end computer to compare improvements. I also tested “Windows only”, as this is the OS our customers are using most. I usually performed all commands 10 times, threw out the best and worst performance times, and calculated the average of the rest. On some tests, like svn log when caching jumps in, I used only the first (much higher) value. I used a locally-installed Apache to circumvent nearly all network latency. Since that is different everywhere and always, I decided to push it out of the equation. I also performed some checkouts on a SSD to rule out HD performance effects.

The Candidate Commands

I was only interested in these three SVN commands:
  • svn checkout
  • svn log
  • svn status
These three commands are the ones Subversion users most often blame for their low performance. Note that svn status operates without a server connection so it is HD-bound and therefore caching is significantly affecting the measurement, which is why I used the first measured value.

Test Machine Specs

  • CPU: 1.6 GHz Dual core Centrino
  • RAM: 2.5 GB
  • HD: 2.5” 120 GB/7200rpm
  • Apache httpd: 2.2.21 (std. Windows build) on localhost
  • SVN 1.6: Subversion 1.6.13
  • SVN 1.7: Subversion 1.7.0

The Test Repository

I used a Polarion repository with a lot of projects and customer-specific configurations for my performance checks. This exemplifies Subversion’s worst case scenario: a huge number of small files in another huge number of folders. As you see I have nearly as many folders as files and the average file size is about 3.2 KB.
  • Repository Size: 32.454.371 Bytes (30.9 MB)
  • Revisions : 2325 revisions
  • Number of Files in HEAD: 5509 files
  • Number of Folders in HEAD: 4267 folders
  • Overall File Size in HEAD: 18.2MB

Test #1: svn checkout

Here are the results of a first test run in the environment I have already described: Graph of test results We then made another checkout on a much faster machine with SSD (Intel core i7-2600, 8GB, Intel 320m 160Gb SSD). These are the results: Graph of another test result You can see the usage of SSD increases the performance dramatically. (Note that server and repo was sitting on a spinning 7200 rpm HD, only working copy was writing on SSD). The ratio is somewhat equal to the slower case above: SVN1.7 is much faster, even with a SVN1.6 server.

Analysis

These numbers are quite impressive! It seems that SVN1.7 is more than twice as fast on an initial checkout even if you are using only the SVN1.7 client. This makes an upgrade very recommendable. If you are using a SVN1.7 Server, the improvements are even higher. The reason for this performance impact is two-fold: 1.   The new working copy format does not need to write so many files / folders The new working copy format holds all SVN data on the topmost folder. No more .svn folders in each directory! As you see on the table, there is no improvement on working copy size though. However, you need many less folders(something around 260 folders max), while we had much, much more folders in old working copy format. Not so extreme, but still impressive, is the filecount: SVN1.7 will store the same files only once. In my Polarion repo are a lot files equal (mainly XML-files), so we end up with not twice of the file count. In Subversion 1.6 the file size roughly quadrupled: SVN working copy comparison graph, SVN 1.6 and SVN 1.7 2.   The new HTTPv2 Protocol uses far less handshakes and so reduces Apache overhead Here is an Apache log of a checkout of a single file (client 1.6 / server 1.6):
127.0.0.1 - - "OPTIONS /repo/.dms/fields HTTP/1.1" 200 143 127.0.0.1 - - "PROPFIND /repo/.dms/fields HTTP/1.1" 207 353 127.0.0.1 - - "PROPFIND /repo/!svn/vcc/default HTTP/1.1" 207 244 127.0.0.1 - - "PROPFIND /repo/!svn/bln/2325 HTTP/1.1" 207 257 127.0.0.1 - - "PROPFIND /repo/.dms/fields HTTP/1.1" 207 353 127.0.0.1 - - "PROPFIND /repo/!svn/vcc/default HTTP/1.1" 207 244 127.0.0.1 - - "PROPFIND /repo/!svn/bln/2325 HTTP/1.1" 207 257 127.0.0.1 - - "PROPFIND /repo/.dms/fields HTTP/1.1" 207 353 127.0.0.1 - - "PROPFIND /repo/!svn/vcc/default HTTP/1.1" 207 257 127.0.0.1 - - "PROPFIND /repo/!svn/bc/2325/.dms/fields HTTP/1.1" 207 359 127.0.0.1 - - "OPTIONS /repo/.dms/fields HTTP/1.1" 200 143 127.0.0.1 - - "PROPFIND /repo/.dms/fields HTTP/1.1" 207 353 127.0.0.1 - - "PROPFIND /repo/.dms/fields HTTP/1.1" 207 353 127.0.0.1 - - "PROPFIND /repo/!svn/vcc/default HTTP/1.1" 207 244 127.0.0.1 - - "PROPFIND /repo/!svn/bln/2325 HTTP/1.1" 207 257 127.0.0.1 - - "PROPFIND /repo/.dms/fields HTTP/1.1" 207 353 127.0.0.1 - - "PROPFIND /repo/!svn/vcc/default HTTP/1.1" 207 244 127.0.0.1 - - "PROPFIND /repo/!svn/bln/2325 HTTP/1.1" 207 257 127.0.0.1 - - "REPORT /repo/!svn/vcc/default HTTP/1.1" 200 1067
Here is the same checkout on a SVN1.7 client: (client 1.7 / server 1.6):
127.0.0.1 - - "OPTIONS /repo/.dms/fields HTTP/1.1" 200 143 127.0.0.1 - - "PROPFIND /repo/.dms/fields HTTP/1.1" 207 353 127.0.0.1 - - "PROPFIND /repo/!svn/vcc/default HTTP/1.1" 207 244 127.0.0.1 - - "PROPFIND /repo/!svn/bln/2325 HTTP/1.1" 207 257 127.0.0.1 - - "PROPFIND /repo/.dms/fields HTTP/1.1" 207 353 127.0.0.1 - - "PROPFIND /repo/!svn/vcc/default HTTP/1.1" 207 257 127.0.0.1 - - "PROPFIND /repo/!svn/bc/2325/.dms/fields HTTP/1.1" 207 359 127.0.0.1 - - "OPTIONS /repo/.dms/fields HTTP/1.1" 200 143 127.0.0.1 - - "PROPFIND /repo/.dms/fields HTTP/1.1" 207 353 127.0.0.1 - - "PROPFIND /repo/.dms/fields HTTP/1.1" 207 353 127.0.0.1 - - "PROPFIND /repo/!svn/vcc/default HTTP/1.1" 207 244 127.0.0.1 - - "PROPFIND /repo/!svn/bln/2325 HTTP/1.1" 207 257 127.0.0.1 - - "REPORT /repo/!svn/vcc/default HTTP/1.1" 200 1067
So even on a 1.6 Server there is a reduced number of http transactions (19 vs. 13), but if you upgrade to a 1,7 Server you got the full power of the new HTTPv2 Protocol(client 1.7 / server 1.7):
127.0.0.1 - - "OPTIONS /repo/.dms/fields HTTP/1.1" 200 143 127.0.0.1 - - "OPTIONS /repo/.dms/fields HTTP/1.1" 200 143 127.0.0.1 - - "PROPFIND /repo/!svn/rvr/2325/.dms/fields HTTP/1.1" 207 359 127.0.0.1 - - "OPTIONS /repo/.dms/fields HTTP/1.1" 200 143 127.0.0.1 - - "OPTIONS /repo/.dms/fields HTTP/1.1" 200 143 127.0.0.1 - - "REPORT /repo/!svn/me HTTP/1.1" 200 1099
Only 6 HTTP requests!

Test #2: show log all

Table of test results from show log, SVN 1.6, SVN 1.7

Analysis

I did not get the reason why logging is slower between a SVN 1.6 client and a SVN 1.7 server. Although this would be something interesting to investigate, I felt it is probably not one of the more common usage scenarios and so didn’t take the time to delve into it. So no big news here: performance is quite the same on my machine, but it seems that the HTTP round-trips have been reduced as well. Log message for SVN 1.6 (client 1.6 / server 1.6):
127.0.0.1 - - "OPTIONS /repo HTTP/1.1" 200 143 127.0.0.1 - - "PROPFIND /repo HTTP/1.1" 207 337 127.0.0.1 - - "PROPFIND /repo/!svn/vcc/default HTTP/1.1" 207 244 127.0.0.1 - - "PROPFIND /repo/!svn/bln/2325 HTTP/1.1" 207 257 127.0.0.1 - - "PROPFIND /repo HTTP/1.1" 207 337 127.0.0.1 - - "PROPFIND /repo/!svn/vcc/default HTTP/1.1" 207 244 127.0.0.1 - - "PROPFIND /repo/!svn/bln/2325 HTTP/1.1" 207 257 127.0.0.1 - - "PROPFIND /repo HTTP/1.1" 207 337 127.0.0.1 - - "PROPFIND /repo/!svn/vcc/default HTTP/1.1" 207 244 127.0.0.1 - - "PROPFIND /repo/!svn/bln/2325 HTTP/1.1" 207 257 127.0.0.1 - - "PROPFIND /repo HTTP/1.1" 207 337 127.0.0.1 - - "PROPFIND /repo/!svn/vcc/default HTTP/1.1" 207 257 127.0.0.1 - - "REPORT /repo/!svn/bc/2325 HTTP/1.1" 200 36418
And here is the SVN 1.7 log (client 1.7 / server 1.7):
127.0.0.1 - - "OPTIONS /repo HTTP/1.1" 200 143 127.0.0.1 - - "OPTIONS /repo HTTP/1.1" 200 143 127.0.0.1 - - "OPTIONS /repo HTTP/1.1" 200 143 127.0.0.1 - - "REPORT /repo/!svn/rvr/2325 HTTP/1.1" 200 3641

Test #3: svn status

Table of test results on svn status, SVN 1.6, SVN 1.7

Analysis

On the local SVN status command, the new working copy structure strikes again: the optimized structure in SVN 1.7 clocks in again more than twice as fast compared to SVN 1.6. Note that I used the whole repository as working copy, and changed around 10 to 20 files deep in the nested folders.

Conclusions

Subversion 1.7 seems to dramatically improve the performance. The new working copy architecture is much faster, and the overall usage is much better: no .svn folders everywhere, so you can easily copy files from your working copy to other places. We can definitely recommend to upgrade at least your clients to SVN 1.7 as soon as possible! However, it’s important to keep in mind that the new client’s working copy format is incompatible with older (SVN1.6 ) clients. So make sure to update all clients you are using in your tool chain (e.g. TortoiseSVN, Subversive, smartSVN, etc.) If some of your clients don’t yet support Subversion 1.7, then you will need to wait to upgrade until they do. EDITOR'S NOTE: ??Polarion will officially support SVN 1.7 server in Polarion 2012, scheduled for release in Q1-2012. However, you can already upgrade your SVN clients like TortoiseSVN to benefit from improved SVN performance.
ABOUT THE AUTHOR Lutz Dornbusch is a senior consultant with Polarion Software’s Professional Services group, which delivers training and consulting services for Subversion as well as the company’s own products. He has more than 20 years’ experience in configuration management and programming, and in his role as Subversion consultant he has successfully migrated various version control systems to Subversion. Lutz also leads the SubTrain Project which delivers open source Subversion training materials.

Comments
by Dreamer
on ‎12-13-2011 12:18 AM
We recently upgraded to Subversion 1.7 and haven't had any issues so far.
by Dreamer
on ‎12-29-2011 04:12 AM
Interesting, I wonder if anyone has compared the performance of the 'SVN COMMIT' between he two versions. This is the one that seems to take the most in my development environment.
by Dreamer
on ‎01-02-2012 07:32 PM
I've upgraded SVN to 1.7 on a test server and have an issue with Polarion – the integrated Repo Browser and SVN Picker don’t work. Support says that “Integration of Subvbersion 1.7 is planned for release Polarion 2012”, therefore the “definitely recommend to upgrade” is maybe not the best choice?!
by Dreamer
on ‎01-02-2012 10:08 PM
@Ron Peled: If you use TortoiseSVN it will perform usually an "svn status" command to get a list of modified files. After this it will calculate the difference which should be sent to the server. The first operation was benchmarked, however the second should be as fast as in previous versions. In my experience, The svn status command is main responsible for slow downs while committing