Up and Down the Network Stack

Web 2.0 Application Acceleration with Memcached

In the last blog I wrote about the importance of implementing 10GbE over 1GbE in your Web 2.0 infrastructure and what features to look for in a 10GbE NIC…since they are not all the same.  In this blog we are going to look at Memcached and our experience in accelerating access to data for improved user experience. 

Memcached is a widely used distributed database caching technology, implemented in tens of thousands of web implementations, including sites such as Facebook, YouTube, Twitter, Salesforce.com, Wikipedia, and Craigslist.  Memcached accelerates dynamic web applications and is now an important part of many social networking infrastructures.

Memcached improves the performance of distributed dynamic web and database applications by caching dynamic object requests and reducing the volume of accesses to the database and storage tiers. This dynamic caching can increase the transaction rates of dynamic web applications and reduce database load. Memcached can either utilize unused memory in application servers or be deployed as a dedicated tier in the database, creating a large, shared pool of virtual object-based cache accessible to all servers.

 
In our Lab and at MyYearbook.com we experienced some additional performance improvements using the Solareflare 10GbE NIC along with its unique OpenOnload application acceleration middleware which reduced the overall overhead associated with Memcached accesses.  The result is increased transaction rates and reduced response times for web applications!

In testing performed with a Solarflare SFN5122F 10GbE server adapter, Memcached response times measured using the Memslap benchmark at 200K transactions/sec were cut in half, from 134μs to 65μs.

Come to Couch Conf on July 29^th in San Francisco and discover how you too can improve Web 2.0 application performance with Solarflare technology.

------------------------------------------------------------------------------
Guest Blogger: Tim Dales, Solarflare Communications

Web 2.0 Application Acceleration

Current Web 2.0 infrastructures were built when 1Gbit Ethernet server adapters were common and inexpensive to deploy.  As the cost of 10Gbit Ethernet continues to decline, network architects can now consider 10GbE in their designs. Using 10GbE in your Web 2.0 infrastructure can unlock bottlenecks, free up valuable resources and save money.  Consider just a few of the benefits:

• 10x performance in your network
• Less cabling = simplified management
•  Reduced power consumption
  - 10x 1GbE = ~ 80W
  - 1x 10GbE = 5W (SFP+)
  - 50 servers will save 3.75 KW
• Frees up bus slots in servers

But all 10Gbit NICs are not the same. Beware of low-priced Intel NICs (AT2) that are built on older Oplin chip technology, which doesn’t operate at PCI Gen 2.0 speeds and has a limited feature set. Additionally, there are a bunch of “converged” 10Gbit Ethernet adapters in the market from storage HBA vendors (QLogic, Emulex, etc.) that are now trying to put all protocols (FC, iSCSI and Ethernet) on the same wire.  Great concept, but costs are high, performance is lagging and simultaneous multi-protocol support is still developing.

So, for 10GbE to accelerate your Web 2.0 applications, the “pure play” is best.  Look for a 10GbE NIC that offers a good out-of-box experience.  What I mean by this is that the NIC goes into the server easily, you install the driver and it just runs, without tuning or fiddling.  Second, make sure your 10GbE NIC offers core scaling features to balance your workloads over all CPU cores and maximize performance.  Third, hardware-accelerated flow steering which reduces the CPU overhead of IP packet transport, providing more CPU cycles to work on the application.

Once you know that the 10GbE NIC has all those features, you will be assured that you are designing a high performance infrastructure to support Web 2.0 application growth and performance. 

In our next blog we will talk about Memcached and how 10GbE can help accelerate this critical part of Web 2.0 application infrastructures.

------------------------------------------------------------------------------
Guest Blogger: Tim Dales, Solarflare Communications

 

Bubba's Number #10

CONGRATULATIONS BUBBA !!!!  10 is a big number. Never thought you'd make it.

If you don't know what I'm talking about, go read

http://mentalburdocks.blogspot.com/2008/11/wheres-head-bubba.html

BTW this is all that's left of my bottle, which was very smooth :)

 

Sex, Lies, and Videotape

Scene from Competitor's Conference Room;  (you'll have to forgive my accent) ...

Yes, Solarflare is the competitor with that amazing OpenOnload product - they're just whipping us out in the street. We need to come up with something. Our stuff is so bad,  it wont even run NetPerf.

I know, lets write our own benchmark that only works with our product and put out a whitepaper rubbishing Solarflare.

Brilliant @"! Yes, and we can make up some really bad results for the Solarflare product!

           ...

... erm, I'd like to give credit to our customers with a bit more intelligence than to fall for that.

But while all publicity is good publicity and it's nice to be recognised as THE ONE TO BEAT! this is all going to get tedious and as tortuous as the plot of Soderbergh's  film .  To prevent such nonsense, we've (with regret) added a note to the IP ownership statement within OpenOnload:

You will not disclose to a third party the results of any performance tests
carried out using OpenOnload or EnterpriseOnload without the prior written
consent Solarflare.

 

So as of openonload-201104-u1 we'll have no more of that, thank you.

In the meantime, if you want some sport, then just ask your favourite competitor rep :

- why do I need to overwrite my system libraries to use your product

- when are you going to support fork()/exec() properly

- when are you going to support multiple processes subscribing to the same multicast groups

- why does the performance of your product degrade so badly when I use a real application

- why did it take me so long and so much pain to get a simple benchmark running

- if you fix the above, then wont you be infringing Solarflare IP, they've been at this for years

On a more positive note, we did make an update release to day to openonload. Contains bug fixes and some performance enhancements.

Enjoy the weekend!

 

 

 

 

 

 

 

You will not disclose to a third party the results of any performance tests
carried out using OpenOnload or EnterpriseOnload without the prior written
consent Solarflare.