Intent: Characterize Performance of Drupal on Solaris 10 vs Linux, without tuning (i.e., 'out of the box' experience)
| Hardware Used | Sun Fire X2100 server; 1x2.2GHz dual-core Opteron; 4GB memory |
|---|---|
| Drupal Version | Drupal 5.3 |
| Drupal Database | 2000 users, 5000 nodes, 10000 comments, 250 vocabulary items; 15 vocabularies (generated with Drupal 'devel/generator' module) |
| Alternative PHP Cache Enabled | PECL APC 3.0.15 |
| Load Generator | Faban 'fhb' 0.9 (check it out!) |
| Nature of Workload | Concurrent hits against font page (not logged in) |
| SAMP Config | LAMP Config | |
|---|---|---|
| Operating System | Solaris 10u4 (within non-global zone) |
SLES 10.1 |
| AMP Stack | Coolstack 1.1: Apache 2.2.3 MySQL 5.0.33 (32-bit) PHP 5.2.0 |
Bundled AMP: Apache 2.2.3 MySQL 5.0.26 (64-bit) PHP 5.1.2 |
Commentary:
- Results at this point are only an indication - shouldn't be used for comparison yet - need more realistic workload
- As Coolstack is optimized for Solaris, and Solaris for Sun Fire servers, results are not totally surprising
- SLES surpasses 1 second avg response time @ around 10 users; Solaris @ around 22 users
- SLES starts throwing '503' errors @ around 10-15 users; Solaris @ around 90-100 users
- In both cases, CPU quickly maxed out (around 85-90% User CPU, 10-15% System CPU) - almost all of this time spent in MySQL
- Minimal disk activity noted
Further Refinements planned:
- Further test results expected in 1-2 weeks
- Test with logged-in users, vary users and pages hit
- Use Coolstack 1.2
- Test against Drupal 6.X
- Use 64-bit MySQL for both
- Use '90% Response time' rather than 'average response time'
- MySQL tuning
Comments (2)
Nov 30, 2007
smattoon says:
Coolstack 1.1 was built with gcc, so optimizations are not as significant as Coo...Coolstack 1.1 was built with gcc, so optimizations are not as significant as Coolstack 1.2, which was built with SunStudio 12. Can't wait to see those results!
Good stuff. What is it about this workload simulation that is not realistic?
Nov 30, 2007
ereid says:
Well, perhaps a better word would be 'incomplete'... I can see scenarios where s...Well, perhaps a better word would be 'incomplete'... I can see scenarios where simultaneous requests against the main page by non-logged-in users ("lots of people suddenly coming to our site 'cuz someone important blogged about it this morning")... but we'll also want to know about how the performance differs with a mix of node requests, and a mix of some logged-in users.