Nexcess Magento Imagine 2014 Performance Breakout
- 1. Simple Steps to Win the Site Performance Battle (and other hopefully interesting stuff) CHRIS WELLS – CEO – NEXCESS.NET LLC SOUTHFIELD, MI USA
- 2. Quick Facts About Detroit, MI • It’s the potato chip capital of the world • based on consumption • It had the 1st mile of concrete road (1909) • Between 6 and 7 mile roads on M-1 (Woodward) • Home to the only floating Post Office in the U.S. • J.W.Westcott Co (Zip code: 48222) • It has the most registered bowlers nationwide • Biggest municipal bankruptcy in the U.S. • Which makes us National Bankruptcy Champions!! • Eat your heart out Jefferson County, AL!! We did 4 x better than you!!
- 3. Notes About This Talk • Lessons learned from MagentoLive Australia • If masses walk out mid-talk I’m being: A)WAY too technical B) Ridiculously boring • Lessons learned from MagentoLive UK • If I sweat profusely: A) DO NOT BEALARMED B) Assume the technical folks located at the back of the room simply swapped out the real talk I had prepared with something they thought was more fun and I’m simply dying a slow 35 minute death on stage before your very eyes.
- 4. The Basics NOT On Today’s Agenda • OS level set up / configuration • Apache or Nginx • Magento caching • Magento EE’s FPC • Varnish /Turpentine or other reverse proxy caching techniques • Other performance extensions • At this point we’re all well aware… • So let’s dig a little deeper into some fun stuff!
- 5. *ON* Today’s Agenda • PHP – 92% • MySQL – 3% (if time permits) • Drives – 5% (if time permits) • If you leave here today with some food for thought then I’m happy!
- 6. PHP – Personal Home Page • Yes, I know.. It’s now “PHP: Hypertext preprocessor” • Magento used to be version locked to PHP <= 5.3 • As of January Magento officially supports PHP 5.4! Woohoo! • You may need to patch your Magento for it to work with 5.4 • So…? • PHP 5.4 is a good deal faster than 5.3 • Empty hash table optimizations • Literal tables • Interned strings • Zend EngineVM tuning • But what does this mean for Magento?
- 7. Notes On Magento • Magento is very CPU heavy • The bulk of this heaviness is within PHP code • MySQL CPU usage is typically bound by PHP’s CPU usage (speaking in single server terms) • Magento is not very I/O heavy (contrary to popular belief) • Again, I/O is bound by PHP’s CPU usage • More on I/O later • Quantitative notes to follow
- 8. Let’s Test Us Some PHP • 1-hour Siege runs (meant to tax PHP, not MySQL) • EE tests • Siege of 20,000 SKUs hitting ~20,000 random URLs • FPC enabled (Redis) • CE tests • Siege of 120 SKUs hitting ~120 URLs. • Bare metal configuration of: • Dell R420 – Dual E5-2420 6-core • H710 512MB NV + 4 x 15K 300GB RAID 10 • 24G RAM • Ok ok, let’s see what we’ve got!
- 9. PHP 5.3.24 (CE Front-end Baseline) 0 20 40 60 80 100 120 140 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 20 25 30 35 40 45 50 55 60 Siegetransacons/second Dura on (in minutes) PHP 5.3.24 ~42 t/sec
- 10. PHP 5.3.24 vs. PHP 5.4.28 (CE Front-end) 0 20 40 60 80 100 120 140 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 20 25 30 35 40 45 50 55 60 Siegetransacons/second Dura on (in minutes) PHP 5.3.24 0 20 40 60 80 100 120 140 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 20 25 30 35 40 45 50 55 60 Siegetransacons/second Dura on (in minutes) PHP 5.3.24 PHP 5.4.28 ~52 t/sec (~23% increase!) Magento cache population
- 11. PHP 5.3.24 (EE Front-end Baseline) 0 100 200 300 400 500 600 700 800 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 20 25 30 35 40 45 50 55 60 Siegetransacons/second Dura on (in minutes) PHP 5.3.24 ~550 t/sec GO GO FPC!
- 12. PHP 5.3.24 vs. PHP 5.4.28 (EE Front-end) 0 100 200 300 400 500 600 700 800 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 20 25 30 35 40 45 50 55 60 Siegetransacons/second Dura on (in minutes) PHP 5.3.24 0 100 200 300 400 500 600 700 800 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 20 25 30 35 40 45 50 55 60 Siegetransacons/second Dura on (in minutes) PHP 5.3.24 PHP 5.4.28 ~615 t/sec (~12% increase!!)FPC / Magento cache population amplifies base PHP version improvements
- 13. PHP 5.3.24 vs. PHP 5.4.28/.14 (EE Front-end) 0 100 200 300 400 500 600 700 800 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 20 25 30 35 40 45 50 55 60 Siegetransacons/second Dura on (in minutes) PHP 5.3.24 PHP 5.4.14 PHP 5.4.28 Even upgrading from 5.4.14 shows gains of ~4%
- 14. PHP 5.3.24 vs. PHP 5.4.28 (Re-index from CLI) 0 10 20 30 40 50 60 70 80 90 100 PHP 5.3.24 PHP 5.3.24 Time In Seconds 86 88 90 92 94 96 98 100 PHP 5.3.24 PHP 5.4.28 ~8% improvement
- 15. Let’s Push Things a Little Further...
- 16. PHP 5.4.28 vs. PHP 5.5.12 (EE/Front-end) 0 100 200 300 400 500 600 700 800 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 20 25 30 35 40 45 50 55 60 Siegetransacons/second Dura on (in minutes) PHP 5.4.28 PHP 5.5.12 ~781 t/sec ~25% better than 5.4 ~41% better than 5.3 WOW FPC / Magento cache population again amplifies base PHP version improvements
- 17. Notes On PHP 5.5.x+ • !! Warning – Magento does not yet officially support 5.5+ !! • APC goes out • Magento configs may need to be tweaked if you’re using APC’s key/val caching • Some extensions will need to be updated / re-coded for sure • OPcache is introduced • OPcache is the name of the bundled ZendOptimizer+ opcode caching system • Seems to work out of the box without too much fuss • More research is needed here – was very surprised with the performance results • !! Warning – Magento does not yet officially support 5.5+ !!
- 18. PHP 5.4.28 vs. PHP 5.5.12 (Re-index from CLI) 85 86 87 88 89 90 91 PHP 5.4.28 PHP 5.5.12 Time In Seconds ~4% improvement
- 19. Let’s Push Things a Little Further…
- 20. PHP 5.5.12 vs. PHP 5.6.0-BETA2 (EE/Front-end) 0 100 200 300 400 500 600 700 800 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 20 25 30 35 40 45 50 55 60 Siegetransacons/second Dura on (in minutes) PHP 5.5.12 PHP 5.6.0 (BETA 2) Essentially equal within margin of error
- 21. PHP 5.5.12 vs. PHP 5.6.0-BETA2 (Re-index from CLI) 72 74 76 78 80 82 84 86 88 PHP 5.5.12 PHP 5.6.0 BETA2 Time In Seconds BUT, we got ~11% improvement on the re-index! ~21.5% better than 5.3.24!
- 22. Let’s…
- 23. PHP 5.5.12 vs. PHP-NG (EE/Front-end)
- 24. Notes on PHP-NG • Removes numerous heap allocations (and de-allocations) • Stores more native data directly on the stack • Removes the need to garbage collect basic primitives (bool, long, etc) • PHP’s reportedWordpress benchmarks show very good results • 26.75 sec -> 18.95 sec (36% improvement) • 9.5M instructions -> 5M internal instructions executed (90% reduction) • Take some comfort in knowing that more gains are on the way from PHP folks directly
- 25. Simple Step #1 – Use PHP 5.4.28 (for now) **But watch out for future PHP releases!
- 26. MySQL – “My SQL” • My is the Michael Widinius’s daughter’s name (this was news to me) • Lots of “brand” options on the MySQL front unlike PHP • MariaDB • Percona Server / PerconaCluster • Oracle’s MySQL • Equally many options on the storage engine front • MyISAM • InnoDB • XtraDB
- 27. Let’s Test Us Some MySQL • ~1 hour re-index runs (ideally meant to put some load on MySQL) • Full CLI-based re-index • Tested latest EE with various versions/brands of MySQL • DB contained 800,000 SKUs (thanks Magento performance team) • Bare metal configuration of: • Dell R420 – Dual E5-2420 6-core • H710 512MB NV + 4 x 15K 300GB RAID 10 • 24G RAM • Ok ok, let’s see what we’ve got!
- 28. Magento CLI Re-index (EE) Time In Seconds 3100 3150 3200 3250 3300 3350 3400 3450 3500 3550 MySQL 5.1.71 Percona Cluster 5.5.34 MariaDB 10.0.10 MySQL 5.6.17 Percona Server 5.6.17
- 29. Magento CLI Re-index (EE) Time In Seconds 0 5 10 15 20 25 Percona Server 5.6.17 MariaDB 10.0.10 MySQL 5.1.71 Percona Cluster 5.5.34 MySQL 5.6.17 Time Spent in "sys"
- 30. Thoughts On The Tests • Re-index runs, while DB intensive, are still PHP heavy • Oracle’s MySQL did surprisingly well • If concurrency were higher I’d expect Percona/MariaDB to excel • DB bound writes (beyond the re-index) were missing • This would tax the DB servers considerably • TCP (and other) overhead for Percona Cluster doesn’t seem to kill performance
- 31. Simple Step #2 – Use a modern version of MySQL (ideally XtraDB based)
- 32. Hard and Solid State Disk Drives • Hard drive vs. disk drive vs. solid state drive vs. SAS, etc. • IOPs (I/O operations / sec) • More is better, but don’t throw money away • All SSDs not created equal • Don’t believe quoted specs! • Don’t buy consumer grade SSDs! • Quality SSDs are ~2x the cost of quality HDD of same size • Know your workload profile!
- 33. Live Client Case Study • Magento EE 1.12.0.1 • ~700,000 SKUs • 350+ orders / hour • 15 x web servers (bare metal) • Dell R420, Dual E5-2450 8-core, 24GB RAM • 2 x DB servers (replicated) – peak @ ~20,000 qps • 1 x file server (shared storage for media) • FPC / Magento cache in Redis • ~1 week sample size
- 34. Database Server IO Requirements Read Write Average IOPs: 3.66 471 Min IOPs: 0.0 6.87 Max IOPs: 830 2250 Standard deviation: 34.5 197.0 95% below: 60.4 795 99.999% below: 151 1310 • Percona Server 5.5.35 • 128GB RAM / 60GB MySQL DB (innodb_buffer_pool = 75G) • 8 x 300GB 15k HDD/SAS RAID 10 - H700/1G RAID card • Replicating to a hot-standby DB server (unused by Magento) ???
- 35. File Server IO Requirements Read Write Average IOPs: 13.7 7.65 Min IOPs: 5.58 2.97 Max IOPs: 85.8 186 Standard deviation: 6.72 11.0 95% below: 24.7 25.8 99.999% below: 42.3 54.6 • Serving /media and /var only over NFS • 24GB RAM - 600GB data on disk • 4 x 600GB 15k HDD/SAS (RAID 10) - H700/512M RAID card
- 36. Notes On Drives • Get quality drives regardless of technology • No consumer SSDs! • Caches are very good at relieving disk-bound I/O • MySQL query cache • Redis/memcaches • RAID card caches • On-disk controller caches • Invest in CPU over IOPs unless you can afford both • Magento is typically bound by CPU, not I/O – Know your workload profile
- 37. Simple Step #3 – Use HDD & SSD drives as needs/costs dictate
- 38. Thank you! ** Thanks for not walking out while I sweat. Come collect your free bag of chips!! CHRIS WELLS – CLWELLS -AT- NEXCESS.NET