Tokyo Tyrant and Tokyo Cabinet are the components for a database used by Mixi (basically a Japanese Facebook). And for work, I got to play with these tools for some research. Installing all this stuff along with the Perl APIs is incredibly easy.
Ultimately I am working on a comparison of Cassandra and Tokyo Cabinet, but I will get to more on Cassandra later.
Ideally the tests I am going to be doing are fairly simple. I am going to be loading a few million rows into a TCT database (which is a table database in TC terms) and then loading key, value pairs into the database. The layout in a hash format is basically going to be as follows:
1 2 3 4 | { "user@example.com" => { "sendDates" => {"2009-09-30"}, }, "123456789" => { "2009-09-30" => {"2287"} }, } |
I ran these tests in the following formats for INSERTing the data into the a table database and as serialized data in a hash database. It is necessary to point out that the load on this machine is the normal load. Therefore it cannot be a true benchmark. Since the conditions are not optimal (but really, when are they ever), take the results with a grain of salt. Also, there is some data munging going on during every iteration to grab the email addresses and other data. All this is being done through the Perl API and Tokyo Tyrant. The machine that this is running on is a Dual Dual Core 2.5GHz Intel Xeon processor with 16G of memory.
For the first round, a few things should be noted:
- The totals referenced below are email address counts add/modified in the db
- I am only using 1 connection to the Tokyo Tyrant DB and it is currently setup to handle 8 threads
- I didn’t do any memory adjustment on startup, so the default (which is marginal) is in use
- I am only using the standard put operations, not putcat, putkeep, or putnr (which I will be using later)
The results of the table database are as follows. It is also worth noting the size of the table is around 410M on disk.
1 2 3 4 5 6 7 8 9 10 | [elubow@db5 db]$ time ./tct_test.pl -b lists/ -D 2009-09-30 -c queue-mail.ini usa: 99,272 top: 3,661,491 Total: 3,760,763 real 291m53.204s user 4m53.557s sys 2m35.604s [root@db5 tmp]# ls -l -rw-r--r-- 1 root root 410798800 Oct 6 23:15 mailings.tct |
The structure for the hash database (seeing as its only key value) is as follows:
1 2 | "user@example.com" => "2009-09-30", "123456789" => "2009-09-30|2287", |
The results of loading the same data into a hash database are as follows. It is also worth noting the size of the table is around 360M on disk. This is significantly smaller than the 410M of the table database containing the same style data.
1 2 3 4 5 6 7 8 9 10 | [elubow@db5 db]$ time ./tch_test.pl -b lists/ -D 2009-09-30 -c queue-mail.ini usa: 99,272 top: 3,661,491 Total: 3,760,763 real 345m29.444s user 2m23.338s sys 2m15.768s [root@db5 tmp]# ls -l -rw-r--r-- 1 root root 359468816 Oct 7 17:50 mailings.tch |
For the second round, I loaded a second days worth of data in to the database. I used the same layouts as above with the following noteworthy items:
- I did a get first prior to the put to decide whether to use put or putcat
- The new data structure is now either “2009-09-30,2009-10-01” or “2009-09-30|1995,2009-10-01|1996”
Results of the hash database test round 2:
1 2 3 4 5 6 7 8 9 10 11 | [elubow@db5 db]$ time ./tch_test.pl -b lists/ -D 2009-10-01 -c queue-mail.ini luxe: 936,911 amex: 599,981 mex: 39,700 Total: 1,576,592 real 177m55.280s user 1m53.289s sys 2m8.606s [elubow@db5 db]$ ls -l -rw-r--r-- 1 root root 461176784 Oct 7 23:44 mailings.tch |
Results of the table database test round 2:
1 2 3 4 5 6 7 8 9 10 11 | [elubow@db5 db]$ time ./tct_test.pl -b lists/ -D 2009-10-01 -c queue-mail.ini luxe: 936,911 amex: 599,981 mex: 39,700 Total: 1,576,592 real 412m19.007s user 4m39.064s sys 2m22.343s [elubow@db5 db]$ ls -l -rw-r--r-- 1 root root 512258816 Oct 8 12:41 mailings.tct |
When it comes down to the final implementation, I will likely be parallelizing the put in some form. I would like to think that a database designed for this sort of thing works best in a concurrent environment (especially considering the default startup value is 8 threads).
It is obvious that when it comes to load times, that the hash database is much faster. Now its time to run some queries and see how this stuff goes down.
So I ran some queries first against the table database. I grabbed a new list of 3.6 million email addresses and iterated over the list, grabbed the record from the table database and counted how many dates (via array value counts) were entered for that email address. I ran the script 4 times and results were as follows. I typically throw out the first run since caching kicks in for the other runs.
1 2 3 4 | Run 1: 10m35.689s Run 2: 5m41.896s Run 3: 5m44.505s Run 4: 5m44.329s |
Doing the same thing for the hash database, I got the following result set:
1 2 3 4 | Run 1: 7m54.292s Run 2: 4m13.467s Run 3: 3m59.302s Run 4: 4m13.277s |
I think the results speak for themselves. A hash database is obviously faster (which is something most of us assumed from the beginning). The rest of time comes form programmatic comparisons like date comparisons in specific slices of the array. Load times can be sped up using concurrency, but given the requirements of the project, the gets have to be done in this sequential fashion.
Now its on to testing Cassandra in a similar fashion for comparison.