Project Mercury Benchmarks: 2000+ Requests Per Second!

joshk's picture

While working through some issues this weekend and preparing another blog post, I finally got around to doing some comprehensive benchmarking on Project Mercury. In the process, I discovered that the first bottleneck I hit running tests from my desktop was the local (last-mile) internet connection, so I switched to running the tests from another EC2 instance. This means that network is not a factor in my results, giving us a real sense of the raw power behind this stack.

For all these tests, I used the Mercury Alpha4 release on a small ec2 instance, loading a staging copy of Mission Bicycles, which is a good "heavy" example in that it has a lot of modules loaded, including Ubercart and Panels. My goal was to measure throughput and response times under various caching configurations, angling for the best results in terms of pages served per second, and delivery time.

I started by cutting things all the way back to nothing, and then added each layer of the caching infrastructure, running benchmarks at each point. The results are quite eye-opening. Can you say 2000+ requests per second? Read on for the full story.

Benchmarks Using ApacheBench on Small EC2 Instances

Step Concurrency Caching Throughput 95% response time Observed Load
1 2 no caching 0.8 req/sec 6880ms 3.5
2 5 page cache 10 req/sec 615ms 4
3 10 apc + page cache 70 req/sec 209ms 6
4 10 apc + aggressive cache 140 req/sec 148ms 10
5 25 apc + cacherouter + aggressive cache 322 req/sec 93ms 12
6 25 mercury 1869 req/sec 37ms 0.1
7 100 mercury 2415 req/sec 83ms 0.1

1. No Caching

As we can see, hosted on a small server, stripped of all caching assistance and loaded down with a heavy module stack, Drupal doesn't perform very well. Less than 1 request per second in throughput sent the load up to 3.5, and page response times were taking more than 6 seconds almost 100% of the time. Thankfully, you'd only run your site under anything approaching these circumstances in pure development mode.

2. Standard Page Cache

Drupal ships with a standard page cache which uses the database to store a fully-rendered page. It doesn't have any bad side-effects, and everyone is encouraged to use it in production. Implementing this common-sense feature got us up to 10 requests per second, although the load was still high.

3. Adding APC

APC allows PHP to run faster and consume less memory by keeping the code compiled and cached in memory. Adding it into the mix allowed me to ramp up the concurrency on the tests to 10, and we saw 70 requests per second in throughput with most of the requests coming back in under 150ms. The box was basically getting swamped, but still handling a lot of requests quite quickly.

4. Aggressive Caching

Drupal's "Aggressive Caching" system cuts back on the amount of processing that needs to occur before/around a pageload, even one from cache. Enabling this basically doubled throughput to 140, and allowed the server to become even more swamped with requests.

5. Cacherouter

One of the most important things to do when running high-performance Drupal, especially in the cloud, is get your Drupal cache system out of the database. APC (or memcached) can act as a high-speed caching backend for full pages, as well as for application variables, menu trees, views, filter output, and more. This is vital to keeping any large site responsive for logged-in users, and allowing Drupal to serve cached pages out of memory (as opposed to from mysql) sped up responses for this test as well.

With this measure in place, we are doing pretty much everything possible to serve as many cached pages as quickly as possible, and we got upwards of 300 requests per second and very good response times, although the box was so loaded it probably would take 30 seconds or more to deliver a non-cached page or respond to any logged-in user.

6. Mercury (Pressflow and Varnish)

Mercury took throughput up by a factor of about 6, delivered the pages faster, and simultaneously reduced load by almost 100x. This is very important, because it means while a few urls are taking a massive pounding — typical results from a link on Digg, Drudge or Perez Hilton — the rest of the system can continue to function. The the truth is there's just no way Apache, PHP and Drupal can compete with Pressflow and Varnish. It's like the difference between a Cessna and an F-15; they're just different classes of tool.

7. Internet-Scale Traffic

Just to see what would happen, I upped the concurrency to 100 and Mercury spat out 50,000 pages in 20.7 seconds for an overall throughput of 2415 requests/second. The average time went down a bit as I think the network interfaces were getting saturated, but this is what we call "internet-scale" traffic. I've included the statpr0n below:

Server Software:        Apache/2.2.11
Server Hostname:        ec2-174-129-144-81.compute-1.amazonaws.com
Server Port:            80

Document Path:          /
Document Length:        12972 bytes

Concurrency Level:      100
Time taken for tests:   20.699 seconds
Complete requests:      50000
Failed requests:        0
Write errors:           0
Total transferred:      673107416 bytes
HTML transferred:       648699608 bytes
Requests per second:    2415.55 [#/sec] (mean)
Time per request:       41.398 [ms] (mean)
Time per request:       0.414 [ms] (mean, across all concurrent requests)
Transfer rate:          31756.33 [Kbytes/sec] received

Connection Times (ms)
              min  mean[+/-sd] median   max
Connect:        2   17 141.9      6    4270
Processing:     3   23  27.2     13     676
Waiting:        2   12  17.4      7     626
Total:          5   40 143.6     21    4279

Percentage of the requests served within a certain time (ms)
  50%     21
  66%     42
  75%     50
  80%     52
  90%     76
  95%     83
  98%    103
  99%    116
 100%   4279 (longest request)

Of course the standard caveats apply: these are only raw benchmarks and don't indicate specific real-world performance. Your mileage may vary.

However, I think walking the caching stack and going from 0.8 to 2415.55 requests per second on the same hardware shows just how much the right stack matters. Literally, compared to an un-configured Drupal install, Mercury handled 3000x as many requests, with 40x less load. Impressive!

Comments

Not just Drupal + Varnish

David Strauss's picture

It should be noted that stages six and seven do not merely add Varnish in front of Drupal. Project Mercury switches out Drupal 6 core for Pressflow 6. Pressflow 6 has extensive differences from standard Drupal 6 designed to support Varnish. Drupal 7 will add support similar to Pressflow 6's.

Yes

joshk's picture

This is true. You can't just stick Varnish on Drupal6. I will update the post to be more clear about that point! (and to fix some grammar)

http://www.chapterthree.com | http://www.outlandishjosh.com

Looking forward for more

meba's picture

Looking forward for more info about D6 + Varnish :-)

Benchmarks really only test varnish

Owen Barton_'s picture

I think it is worth noting that much of the other performance tweaks and caching layers in Mercury are made pretty redundant in this particular benchmark (although of course would all add value in most real site patterns), because Pressflow+Varnish allows pages to be served completely from the Varnish cache (i.e. RAM). Hence, because there are no authenticated users (or cache invalidation) with the ab benchmarks neither Apache nor Drupal is involved at all beyond the first request - i.e. a similarly sized page served by Plone, Wordpress or static HTML should get near identical results.

Indeed

joshk's picture

...neither Apache nor Drupal is involved at all beyond the first request - i.e. a similarly sized page served by Plone, Wordpress or static HTML should get near identical results.

This is a valid point, but what makes these results interesting is precisely the origin point. Drupal's power as a CMS lets people run sites that just aren't possible in WP, or could require too much specialized knowledge to build in Plone, and would be nightmarish to maintain as static files. With Pressflow, you can use Drupal with Varnish, and that's a big win.

Also, the other enhancements on the stack are still germane: they help us insure that during this kind of barrage, the rest of the site remains functional/snappy.

It's true that the top-line number here is a raw power benchmark of Varnish, but knowing that these are the numbers you can hit, and that with server load staying low you can still generate other unique pages on a good timetable is pretty important IMHO.

This was also my original

Dries's picture

This was also my original reaction, to be honest. The benchmarks really only test Varnish, and no longer test Drupal, PHP, MySQL or Apache. That is still GREAT because in some cases, one's website can be cached extensively with Varnish. In many cases it simply can't and in that case we're back to serving slow non-cached pages at 1 second/page (or whatnot). In those cases, Varnish still helps, but if you do the math, non-cached pages will be what you are concerned about. Here, the "real" solution is probably horizontal scaling with memcached (or vertical scaling) and Varnish will start to matter less compared to the ideal scenario used in these benchmarks. Nothing new or shocking here, but important to keep in mind when looking at the data. At the end of the day, all performance improvements help, and clearly, this is a significant one.

For authenticated user

moshe weitzman's picture

For authenticated user caching, I have been doing some D7 work on a drupal_render() cache . Am blocked by a couple of thorny DBTNG issues.

An eventual, long-term goal

David Strauss's picture

An eventual, long-term goal ought to be serving almost all pages from edge caches -- even to authenticated users -- and simply altering or including the user-variant content as necessary. There's nothing sacred about serving only anonymous users from the edge cache; it's just hard to do otherwise.

Nice writeup. The narrative

christefano's picture

Nice writeup. The narrative was especially helpful.


Exaltation of Larks
Founder, CEO
http://www.larks.la  
Droplabs
Founder, Lead Burrito Analyst
http://droplabs.net  
Greater Los Angeles Drupal
Organizer, Drupal Adventure Guide
http://drupal.la  

Comments

kbahey's picture

Josh

Note that aggressive caching is now removed from D7. So perhaps taking out aggressive caching would be appropriate for future benchmarks.

David

Is the Pressflow cache header patch based on this 2bits article, or on a backport of D7's changes to cache headers?

Drupal performance tuning, development, customization and consulting: 2bits.com, Inc..
Personal blog: Baheyeldin.com.

Drupal performance tuning, development, customization and consulting: 2bits.com, Inc..
Personal blog: Baheyeldin.com.

Khalid, Aggressive page

Damien Tournoud's picture

Khalid,

Aggressive page caching has not been removed from core 7, it was merely cleaned-up and renamed. The ability not to call hook_boot() and hook_exit() is still there ($conf['page_cache_invoke_hooks'] = FALSE), and necessary for Varnish to be able to cache pages for anonymous users.

Damien Tournoud
http://drupalfr.org

Damien Tournoud

Pressflow uses the same

David Strauss's picture

Pressflow uses the same system as Drupal 7's lazy session creation and (cache) header management. This is a big departure from Drupal 6's code, not a minor patch. There were several cycles of updating the D7 patch and updating PF5/6. We are currently running Pressflow's session and header code on Drupal.org. We're waiting on a better configuration for Squid or moving to Varnish before we take full advantage of it, though.

Well done!

omega8cc's picture

My private obsession on making Drupal sites running fast was greatly inspired by many performance related patches for 5.x and by many great 2bits.com articles, and now the credit goes also to people behind Pressflow and Mercury project. Just a day before Pressflow were opened to download by anyone, we wrote to David Strauss we are testing and will post results of using Ncache on the front and Nginx as a backend. We are collecting now some real world usage examples to get a good base for a benchmark similar to posted here by Josh. Thanks guys!

~Grace -- Turnkey Drupal Hosting on Steroids -- http://omega8.cc

Would love to see us do some

Dries's picture

Would love to see us do some tests with Drupal 7 to identify what patches we need (and can) move from Pressflow to Drupal 7. We already did a lot of work to make Varnish work out of the box with Drupal 7 but there is more we can do ... we have to make sure that Drupal 7 scales, and that it is at least as fast as Drupal 6.

I don't believe Drupal 7's

David Strauss's picture

I don't believe Drupal 7's Varnish support is in any way less polished than Pressflow's. If anything, there are outstanding Drupal 7 changes that still need backporting to Pressflow.

Catch's Patch

joshk's picture

When it comes to running high-performance sites, I think the single most important patch out there is this one. Adding reverse proxy (and CDN) can extend performance for cached/cacheable pages, but those pages still need to be generated by Drupal. Allowing for a non-datbase object-level cache will speed that up quite a bit, as well as making it possible to build much faster sites for logged-in users.

There's a bit of a tension between how good Drupal does as a simple "out of the box" tool deployed on low-end shared hosting, and how many tools there are in its system for running in high performance environments, which require root access on at least one server, and more intimate knowledge of other parts of the stack.

I tend to agree with Catch's sentiment here:

if we can easily install node_cache, user_cache and term_cache.module in contrib with no core hacks due to this patch, then IMO a very small penalty in core would be acceptable trade-off for much bigger contrib gains.

I think as we move more good features into core, it's only natural that it will be a bit slower, at least at first. Drupal 7's core is like Drupal 6 + CCK, so there's a non-trivial difference in complexity right there. Perhaps beyond the code freeze, we can keep an open mind about performance-enhancing patches that don't alter any APIs?

non-cached pages on pressflow too?

catch's picture

It's not clear from the numbers whether the uncached pages were done on pressflow as well or vanilla Drupal core? Any chance of a comparison there?

It was Drupal core

joshk's picture

Though I don't expect a huge amount of variance from Pressflow. This isn't making use of any other advantages (e.g. mysql replication, etc) and the actual code differences to run Varnish-friendly headers, etc, are small. Shouldn't really impact performance much at all.

There is internal work at

David Strauss's picture

There is internal work at Four Kitchens to deliver most page components through ESI, which should considerably accelerate pages for even authenticated users.

Ooooooh!

joshk's picture

That would be like block caching on steroids.

Yep, I'm building it

David Strauss's picture

Yep, I'm building it directly into the block caching system. Blocks are the perfect units for integrating with ESI. The extra win is that pages can have long cache lives even if some blocks (like the "latest forum posts") don't.

ESI? Can you fill me in on

davidseth's picture

ESI? Can you fill me in on what that is? Sorry if it is a basic question...

Edge-side includes:

Exactamundo!

joshk's picture

The extra win is that pages can have long cache lives even if some blocks (like the "latest forum posts") don't.

Exactly that lets you keep static-y pages around for a long time, and also gets you out of a bunch of active-cache-invalidation pickles (potentially).

I'm hoping we can get something equivalent to the block cache system into Panels as well, and integrate that the same way.

Ajaxify Regions

mikeytown2's picture

New module that takes the pain out of using authcache, http://drupal.org/project/ajaxify_regions. Using this one could use boost for limited support of an authenticated page cache. This could merge boost & authcache into one. It's also very applicable to Edge-side includes as the Ajaxify Regions module already slices up the page. It seems like we are starting to converge towards a page cache with the "extras" retrieved in a second request.

WoW!

omega8cc's picture

With every comment it sounds better and better! Now this thread/group is my start-page =)

~Grace

Hot graphs

joshk's picture

I did a graph to help promote this on my blog. Came out looking HOT!

Could you please share with

Garrett Albright's picture

Could you please share with us what flags you used for ApacheBench when you got these numbers?

I'm doing some experimenting on my own server just to see how it compares. I've never really played with Apache Bench before, and I'm noticing that, for example, my "Requests per second" value increases by over 100% if I use AB's "-k" flag to use KeepAlive connections, without making any changes on the server. Clearly the flags used when AB is called are significant.

Don't use keepalive

joshk's picture

You probably don't want to user keepalive in these tests, as that causes each concurrent thread to maintain its connection, which is not the case in real-world traffic results. The simple syntax I used is:

ab -c {concurrency level} -t 60 http://domain.com

This is a bare-bones/brute-force benchmarking command. Do NOT run at -c 100 (my last result) unless you know what you are doing, and NEVER run benchmarks against production servers. This is an easy way to accidentally crash a box.

Please use bechmarking tools with caution. Hitting an un-tuned server with 100 concurrent threads from apachebench can easily send the system into swap, , especially on a server-to-server connection where networking wont impose a de-facto cap. Id this happens, you may need to wait several minutes (or hard reboot) to recover. Start with one thread and ramp-up from there.

You probably don't want to

Garrett Albright's picture

You probably don't want to user keepalive in these tests, as that causes each concurrent thread to maintain its connection, which is not the case in real-world traffic results.

Do you think so? Some quick research reveals this Wikipedia article which implies that the major browsers have supported persistent connections for some time, and this one states "In HTTP 1.1 all connections are considered persistent, unless declared otherwise."

The objective is to

alanburke's picture

The objective is to replicate multiple users grabbing a single page simultaneously, rather than a single user grabbing the same page repeatedly.
Hence the need to treat each request as separate.

Alan

Well, ideally, we'd be

Garrett Albright's picture

Well, ideally, we'd be simulating a user downloading a page, plus all linked CSS and JS files and embedded images and so on using KeepAlive, just as a web browser would. Apache Bench is only downloading the page, then, and not the linked/embedded files?

Maybe I jumped to conclusions on how AB works. But docs on Apache's site are sparse, and it doesn't seem to have a man page on my machine…

EDIT: Scratch that; it has a man page. (I typed "man ab" in the wrong terminal tab previously…) However, the man page is exactly the same as the page linked above… Little info on how it actually works.

And ideally...

joshk's picture

If you're running this in the kind of environment where 2000 req/sec is a going concern, nothing but the drupal page would be coming off this layer. Everything else would be in a CDN. ;)

There are other tools we can use here. Jmeter allows you to easily fetch all related files from an HTML page, for instance. I am using ab in this case because running a jmeter test from my desktop maxes out my local DSL well before we start testing the virtue of the Mercury stack.

Additionally, static content

David Strauss's picture

Additionally, static content would be trivially cached in Varnish as well, even with no CDN.

Test PressFlow on Acquia Hosting?

Dries's picture

I'd love to take PressFlow for a spin on Acquia Hosting to see how it scales horizontally on AWS. Something we could experiment with.

Definitely

joshk's picture

I've been talking a bit about this w/Kieran and Chuck. I think it can help a lot w/the Fields project, as putting this in on a shared cluster type system would be a really big win.

One of the near-term todos is getting at 64bit version of this out, which is important for doing these in higher-end production environments, but please feel free to run w/it.

Durr, PressFlow, not Mercury

joshk's picture

Of course, you can run PressFlow right now. I'm sure the engineering squad over there is more than capable of configuring Varnish on their own. :P

I discussed it with my team

Dries's picture

I discussed it with my team at Acquia and we'll work towards adding PressFlow to our list of distributions/applications to run continuous benchmarks with. We probably won't be able to work on this until after DrupalCon, but we'll report back as soon we have some data to share.

Hello Josh, What kind of

giorgio79's picture

Hello Josh,

What kind of security levels will Mercury have? Recently one of my non drupal sites was hammered with a DDOS attack, and that was when I learned how nasty they can be.

Reading up on Amazon AWS I notice you should handle this in your application layer:
http://developer.amazonwebservices.com/connect/message.jspa?messageID=13...
http://developer.amazonwebservices.com/connect/message.jspa?messageID=13...

****Me and Drupal :)****
Clickbank IPN - Sell online or create a membership site with the largest affiliate network!
Review Critical - One of my sites

DDOS is rough

joshk's picture

It's very difficult to deal with a true DDoS attack. If initiated from a botnet, it's virtually impossible to distinguish the DDoS from "real" traffic. I"ve spoken with the aiCache folks a bit, but have yet to use their tool on any live sites yet. They have some benchmarks which show even higher performance than Varnish, and if they offer nice tuning options for DOS situations, that would be another thing that justifies their commercial price tag.

Nice article. I'm just about

mrfelton's picture

Nice article. I'm just about to start running benchmarks against one of my sites and I'm just trying to get to grips with tools like ab and httperf. I'm interested to know why you increased the concurrency in your ab command each time you ran a different test? I would have thought that you would keep the test conditions the same for each step of the process, otherwise, how can you compare one set of results against the other?

--
Tom
www.systemseed.com - drupal development. drupal training. drupal support.

Increasing Concurrency

joshk's picture

I did start by running the same concurrency levels, but was looking to "max out" each stack configuration. Basically, the -c flag for ab is a rough metric for tuning how hard/fast it benchmarks. The critical metric for knowing how well your stack will hold up is requests/second.

This is also the critical metric to try and determine from your existing traffic stats. Typically site-owners talk in "visits per month" which is unfortunately almost meaningless when trying to assess performance. What you need to try and figure out is your peak throughput. A site may have 10M visits a month, but the odds that they're evenly distributed through that time are slim-to-nil.

Care to share your jmx?

JacobSingh's picture

We (at Acquia) are about to start some D7 benchmarking, and would love a leg up by looking at your jmx.

Did you handle authenticated users in these tests? What about writes?

Get in touch :)
-Jacob

It's ab!

joshk's picture

It's apache bench at the moment, but let's collaborate on this stuff.

Im sorry, but what is Project

Rockland Steel's picture

Im sorry, but what is Project Mercury?

Im failed searching about it in google

Project Mercury

got it. thanks

Rockland Steel's picture

got it. thanks

Using Google

JohnChapman's picture

Simply using two words in google search:

drupal mercury

returned quite a number of pages. This is a very simple search. I'd recommend additionally that you beef up your search skills, as searching successfully is invaluable.

something i want to add one

bennos's picture

something i want to add
one of my drupal projects.

Normal Drupal not pressflow
APC as Opcodecache, Memcache for Page caching(authored users) and boost for anonymus users.
Webserver is Apache 2.2 and MySQL is 5.1 in Percona Edition (nearly all tables are innodb tables).

Concurrency Level: 100
Time taken for tests: 35.713 seconds
Complete requests: 50000
Failed requests: 0
Write errors: 0
Total transferred: 1517850000 bytes
HTML transferred: 1498250000 bytes
Requests per second: 1400.05 [#/sec] (mean)
Time per request: 71.426 [ms] (mean)
Time per request: 0.714 [ms] (mean, across all concurrent requests)
Transfer rate: 41505.31 [Kbytes/sec] received

The frontpage has a some articles and a views tagcloud.
The test above are more tests with pressflow and varnish. A normal drupal installation can also be tuned into high performance.

My Drupal Job exchange:
Jobs Düsseldorf

osu_bucks's picture

quick question, what's the best practice when we run the ab command, do we run it from the server or from machine elsewhere? This is mine running Linode 360 with nginx, xcache, pressflow and boost from the server. What is non-2xx responses? Is it bad?

Concurrency Level: 100
Time taken for tests: 0.875 seconds
Complete requests: 10000
Failed requests: 0
Write errors: 0
Non-2xx responses: 10028
Total transferred: 3890864 bytes
HTML transferred: 1855180 bytes
Requests per second: 11430.91 [#/sec] (mean)
Time per request: 8.748 [ms] (mean)
Time per request: 0.087 [ms] (mean, across all concurrent requests)
Transfer rate: 4343.37 [Kbytes/sec] received

Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 1 1.0 1 7
Processing: 2 7 1.1 7 14
Waiting: 1 3 0.9 3 12
Total: 2 9 1.6 9 14

Percentage of the requests served within a certain time (ms)
50% 9
66% 9
75% 10
80% 10
90% 11
95% 11
98% 12
99% 12
100% 14 (longest request)

ab

mikeytown2's picture

Be sure to set gzip headers since the vast majority of browsers accept gzip.
ab -c 100 -n 10000 -H 'Accept-Encoding: gzip,deflate'

look here

bennos's picture

look here http://en.wikipedia.org/wiki/List_of_HTTP_status_codes

everthing that is not status 200 could be a problem.

Run AB on the server. On the server you have no limitation of the network.

HTML transferred: 1855180

Jamie Holly's picture

HTML transferred: 1855180 bytes / Complete requests: 10000 = 185 bytes per page. You got a problem. There's a chance that all your requests are sending a redirect.

My suggestion would be to rerun the test like this:

ab -n 1 -c 1 -v 2 {url}

That will print out the HTML returned from the page above the test results.


HollyIT - Grab the Netbeans Drupal Development Tool at GitHub.

Wow!

kpdirection's picture

Wow! Love the results.

I have two dedicated servers for different clients. One runs ab at 14.5 requests per second, and the other runs at 75 requests per second. Even 500 would be wonderful for me and the clients!

The thing I find with benchmarking is that there has to be as many 'knowns' in the equation as possible. I get 75 requests per second with my site, but if I set up a standard 'benchmarking' site on there and measured it, together with knowledge of other servers, I'd be able to see if it was my implementation that was at fault or the server setup. (yes, I know at my scores it's probably the server setup...)

I know I've a lot to learn about configuring memcache and nginx, and the correct combinations, but can anyone provide exact setup details (or a 'benchmark' installation of Drupal) that I can use so that I can take the actual Drupal site out of the equation? Does that already exist somewhere?

Regards,

Pete.

While the number is

jcisio's picture

While the number is wonderful, it's far from real world. When I do a test on a production server with a single SATA2 HDD, I get this:

[~]# ab -c 50 -t 10 http://www.mysite.com/
This is ApacheBench, Version 2.0.40-dev <$Revision: 1.146 $> apache-2.0

... bla bla ...

Server Software:        Apache/2.2.14
Server Port:            80

Document Path:          /
Document Length:        77836 bytes

Concurrency Level:      50
Time taken for tests:   10.971 seconds
Complete requests:      10773
Failed requests:        0
Write errors:           0
Total transferred:      844467867 bytes
HTML transferred:       839257263 bytes
Requests per second:    1077.20 [#/sec] (mean)
Time per request:       46.417 [ms] (mean)
Time per request:       0.928 [ms] (mean, across all concurrent requests)
Transfer rate:          82459.49 [Kbytes/sec] received

So, 1000+ rps on a cheap and in production server, only with Apache + boost.module. The homepage is full of Panels (dozens of mini panels) and has more than 30 Views. But with the static cache, like Varnish, there is no PHP or MySQL at all.

In the real scenario, that server can hardly serve 5 pageviews/second with the same configuration (Apache, memcache.module, boost.module and page cache enabled). Not only are there way too many static requests but also not all Drupal requests are cached, particularly when a large part of visitors is from Google.

@kpdirection: for your site, the benchmark results seem normal, be not afraid. If you have 10 req/s in the benchmark, that means 36000 pageviews hourly or nearly 1 million pageviews daily! However, to optimize, the simplest one is boost.module, especially when you have little content but too much traffic (so that it caches well).

html transfered bytes size

osu_bucks's picture

I see why my HTML transferred was 1855180 bytes, because I was doing ab on http://mydomain.com/. It is a redirect. Here's the ab command going to http://www.mydomain.com/ instead, and this time it doesn't have non-2xx responses. When I perform ab with n1 c1 -v 2 and -h gzip, deflate, it returns a portion of my html... I'm going to read up and see if that's by designed or suppose to return full html page.

Server Software: nginx/0.7.65
Server Port: 80

Document Path: /
Document Length: 11518 bytes

Concurrency Level: 100
Time taken for tests: 0.866 seconds
Complete requests: 10000
Failed requests: 0
Write errors: 0
Total transferred: 118960000 bytes
HTML transferred: 115180000 bytes
Requests per second: 11552.96 [#/sec] (mean)
Time per request: 8.656 [ms] (mean)
Time per request: 0.087 [ms] (mean, across all concurrent requests)
Transfer rate: 134212.91 [Kbytes/sec] received

Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 0 0.5 0 6
Processing: 2 8 0.9 8 15
Waiting: 1 7 2.0 8 13
Total: 3 9 0.8 9 16

Percentage of the requests served within a certain time (ms)
50% 9
66% 9
75% 9
80% 9
90% 9
95% 10
98% 10
99% 11
100% 16 (longest request)

3000 request per second?!?! Please help!

blue92877's picture

Hi, I'm Blue, and I am a Windows user ( monotone group chant ... hi blue )

Like many, my I came across this post and many others while researching how to make my Drupal site faster. I've only developed on a Windows machine (WAMP).

At this juncture, to be honest, in my Drupal development life, I've been all too ready to just give up on Drupal simply that its WAY TOO DAMN SLOW. I came across this article, researched about caching engines, mercury, pressflow, and thought, OK, I'll give it a try. Looks like everyone is doing this on a Linux box, so, I'll install an instance of Drupal on Linux.

And I did, and I'm down right flabergastedly amazed and confused as to the results.

Last night I installed a copy of UBUNTU 10.0.4 Desktop.

I used apt-get to install Apache2, PHP5.3, and MySQL. In created a vanilla instance of Drupal 6.19 under /var/www/drupal and installed it via firefox - nothing out of the ordinary. I have about 40 modules installed and enabled (cck, views, path auto, etc).

Let me be clear: This is a default installation. I also used Devel to create 5000 nodes of various content types (blog, forum, page, story, etc), 50 users, and 10 comments per node

  • NO CACHING at all, period
  • NO APC
  • NO MEMCACHE
  • NO MEMCACHED
  • NO PRESSFLOW
  • NO BOOST

Here are the results:

ab -n 1000 -c 100 http://localhost/drupal

Concurrency Level:      100
Time taken for tests:   0.357 seconds
Complete requests:      1000
Failed requests:        0
Write errors:           0
Non-2xx responses:      1004
Total transferred:      562240 bytes
HTML transferred:       311240 bytes
Requests per second:    2801.45 [#/sec] (mean)
Time per request:       35.696 [ms] (mean)
Time per request:       0.357 [ms] (mean, across all concurrent requests)
Transfer rate:          1538.17 [Kbytes/sec] received

Connection Times (ms)
              min  mean[+/-sd] median   max
Connect:        0    1   1.6      1       8
Processing:     5   32   8.2     31      49
Waiting:        3   31   8.0     30      47
Total:         10   34   7.5     33      50

Percentage of the requests served within a certain time (ms)
  50%     33
  66%     38
  75%     42
  80%     42
  90%     43
  95%     44
  98%     45
  99%     46
100%     50 (longest request)

Can someone explain what's going on here? What am I missing? This is literally 1000 x faster (yes, 3 zeros) than the exact same setup on my windows machine. Is linux that much more efficient? Is something installed during apt-get that I'm not aware of that's caching stuff for me? On windows it takes my admin page 18 seconds to load, with 16 of that latency. On Ubuntu, I load the exact same config in less than 2 seconds, same machine. Which, btw, let me give you my machine specs - Acer Timeline 4810T Core 2 Solo 1.4GHZ 4GB Ram.

Any thoughts?

I had to switch to windows (I'm dual booting.) I wanted to add my windows results. It's the same setup EXCEPT I have Xcache installed and enabled.

Server Software:        Apache/2.2.15
Server Hostname:        localhost
Server Port:            80

Document Path:          /gastonia619
Document Length:        237 bytes

Concurrency Level:      100
Time taken for tests:   8.367 seconds
Complete requests:      1000
Failed requests:        0
Write errors:           0
Non-2xx responses:      1000
Total transferred:      476000 bytes
HTML transferred:       237000 bytes
Requests per second:    119.51 [#/sec] (mean)
Time per request:       836.748 [ms] (mean)
Time per request:       8.367 [ms] (mean, across all concurrent requests)
Transfer rate:          55.55 [Kbytes/sec] received

Connection Times (ms)
              min  mean[+/-sd] median   max
Connect:        1    1   0.7      1      18
Processing:   223  795 116.8    825     909
Waiting:      221  791 119.0    822     907
Total:        230  796 116.7    826     910

Percentage of the requests served within a certain time (ms)
  50%    826
  66%    842
  75%    845
  80%    846
  90%    888
  95%    895
  98%    909
  99%    909
100%    910 (longest request)

before adding xcache, my req/sec were 1.5, average.

Non-2xx responses:     

brianmercer's picture

Non-2xx responses:      1004

and
Document Length:        237 bytes

You're benchmarking 404s or something.
Do a "curl http://localhost/drupal" and find out what you're getting.

301, not 404

bibo's picture

You're benchmarking 404s or something.

It's the "something", and that is a 301 redirect caused by Apache / .htaccess -rules that forward:
"http://localhost/drupal" ==> "http://localhost/drupal/"

Do a "curl http://localhost/drupal" and find out what you're getting.

Or just add the missing "/" to the end of the request to get real results.

And then add Varnish to the stack and marvel again at the 3k-10k requests per second. This time with the real page ;)

High performance

Group notifications

This group offers an RSS feed. Or subscribe to these personalized, sitewide feeds:

Hot content this week