(gs) Update: Improved Web Response Times
May 20th, 2011 at 12:14 pm
At the beginning of the year we mentioned several initiatives we’ve been working on to improve overall (gs) performance. Last month we told you about email performance improvements. Today we’re happy to tell you about ongoing storage improvements we’ve made that have improved overall web response times across the Grid.
For the last couple months we’ve been quietly enhancing the (gs) storage infrastructure with the addition of a caching layer built with extremely high speed Solid State drives. We’re about 80% complete with this work and happy to tell you that Grid Clusters with this new enhancement are reporting 25% speed increase for average web response.
As you can see above, storage segments using SSDs are responding faster. The graph shown here compares web response times from three (gs) storage segments with SSDs to three without. Before installation, trial storage performed roughly the same as other segments. As SSDs were added, response times decreased and settled around 25% faster than the rest.
This is a notable improvement, but just one enhancement in ongoing improvements we’re continuing to implement across the (gs) storage architecture.
What is a Solid State Drive?
Put simply, it’s a storage drive that has no moving parts. No spinning disks mean significantly faster data access times, though they generally can’t hold as much data as traditional disk. The speediness of SSD’s make them a perfect fit for a caching layer, where speed is of the utmost importance.
Why is Caching Important?
Good caching tends to make a big difference in a large, clustered environments like the (gs), because the most ‘popular’ data is what tends to slow down the traditional disks leading to increased response times. With a fast cache layer in place, the popular data comes directly from cache, taking the load off the rest of the storage system which leads to overall speed increases for all users.
Does this apply to my site yet?
Right now all customers on clusters 3, 4, 7 ,8 and 9 are using SSDs and work has started on clusters 1 and 6, with 2 and 5 scheduled immediately after. Upgrades are scheduled for upcoming normal maintenance windows to make the transition as convenient as possible. To find out which cluster your (gs) resides on, log in to the AccountCenter and click on your Server Guide from your grid control panel.
Ongoing Improvements
We’re happy with the results from SSDs and the Dovecot email rollout , but we’re not stopping there. We’ve recently started increasing the physical RAM across all (gs) nodes, and we’ll be reporting on that as it gets closer to completion.
Thanks for reading! Feel free to leave any comments below and join the discussion in the User Forums.
» 
May 20th, 2011 at 12:29 pm
Your counting is quite peculiar. I’m on gs cluster 2 — have been for years. Why are you skipping it?
May 20th, 2011 at 12:44 pm
Sorry if this was unclear – Cluster 2 upgrade is underway and about 25% done.
May 20th, 2011 at 1:04 pm
Excellent improvement! My site actually is a bit faster- this is great
It’s reasons like this why I wouldn’t choose any other company to host my sites! Keep it rollin’!
May 20th, 2011 at 1:31 pm
Excellent news. I’m running several sites on the “grid” and I always feel like I’m in good hands with Media Temple.
May 20th, 2011 at 1:45 pm
Thanks Matt!
May 20th, 2011 at 1:45 pm
Uh, hello? Anyone notice how clusters 2 and 5 aren’t even mentioned? Why? 3, 4, 7, 8 and 9 are done, 1 and 6 scheduled but 2 and 5 forgotten?
May 20th, 2011 at 2:22 pm
Apologies for the confusion – cluster 2 is underway (25%) and 5 is scheduled for immediately after.
May 20th, 2011 at 8:24 pm
For SSD that is 100s times faster in some aspect then a HDD, 25% improvement isn’t a lot. Especially factor in the cost of an SSD compare to HDD.
Why is that?
And … where is (cs)
May 20th, 2011 at 9:36 pm
I’ve got sites on cluster 01 and my monitoring shows that the response time got substantially better on April 20th.
May 21st, 2011 at 10:08 am
@Edward, the data shown includes non-cached responses as well. If we showed just SSD response, it would be much more drastic a change, but only applicable to cached files. This shows the overall performance change that customers are seeing – in our opinion a slightly more honest view of the improvement.
May 22nd, 2011 at 11:44 am
I recently moved my website to MediaTemple and I’m very pleased with it.
It’s been 48 hours and there is not issues so far.
My website is opening 5.5 times faster than before
May 22nd, 2011 at 12:57 pm
Nice to see improvements being made to (gs)
Next up – database latency optimizations… I hope;-)
May 22nd, 2011 at 6:24 pm
I’m not even a customer of (mt), but I have a suggestion:
Try out Percona DB server instead of MySQL. It’s the same exact thing, except it’s better performing. It might help you guys a lot. Check it out.
May 23rd, 2011 at 9:32 am
Great idea Adam – we have been working closely with Percona and are already using some of their tools to improve database insight and performance.
May 30th, 2011 at 6:31 am
This is good news. I’m a bit mystified why your oldest customers go next to last (I’m on GS01). This is almost always the case with GS upgrades that 01 is at the end.
RE @Oliver’s question, when are database latency optimizations scheduled?
May 31st, 2011 at 10:44 am
Thanks for the reply, Frank. Changes are normally made to later clusters first because the risk to customers is lower. After the initial rollout though, we should be able to get cluster 1 higher in the queue. I’ll bring it up when we schedule the next improvements.
Re: databases, Percona tools are helping us tweak the pool settings and we’ll try to publish more information as it solidifies.
June 2nd, 2011 at 12:01 pm
Agree with Oliver, “Next up – database latency optimizations… I hope
”
I’m on GS 5, when is the anticipated date of completion? I was planning to move host come renewal time in a few months due to the latency but ‘may’ stick around. Would be interested to know plans for mysql performance improvements as well!
June 2nd, 2011 at 12:09 pm
This is great. I really like the MT, and now is even better. Thanks!
June 2nd, 2011 at 2:44 pm
@Trev,
Thanks for commenting! We’re prepping cluster 5 right now with data migrations and expect the SSD installs to begin at the end of next week. The whole process should take a week and half or so because we’re cycling storage during normal maintenance windows.
Re: databases, we’d love to get more detail on your experience. Do you mind heading over to the forums to chat?
June 2nd, 2011 at 3:44 pm
Can you please provide more information on the criteria you use to determine if content gets cached or not? Does the type of content matter (HTML, image, PHP)? Does the type of CMS matter? I’m using WordPress with W3 Cache already. How would this be affected for my high-traffic web sites.
Also, how often does the cache get refreshed and how does it know if the content of a web site changes?
Thanks!
June 2nd, 2011 at 9:33 pm
Thanks for the info Matt,
Whats the progress on cluster 6 and when the transition is expected to be completed?
TTFB has been very during last couple of months, thanks If you have improving your infrastructure.
Could you explain lit bit more about your caching methods or technologies? OR are you also going to add some opcode or object caching systems (like memcache) etc to save your bandwidth and better performance.
June 3rd, 2011 at 6:11 am
Your transparency to the customers is the main reason I love hosting with you guys. Even when their are problems you guys are up front with everything. Never let that change.
June 3rd, 2011 at 10:03 am
I noticed some speed improvement but still had only 98% uptime in May.
June 3rd, 2011 at 10:18 am
@Aaron, the new SSD layer is part of our ZFS deployment so it uses the Solaris flavor of Adaptive Replacement Cache (ARC). We aren’t intentionally targeting a specific CMS, but the way an app uses the filesystem will definitely affect what ARC sees.
June 3rd, 2011 at 11:45 am
Thanks Dan – let’s follow up to find where our uptime measurements are different.
June 3rd, 2011 at 12:21 pm
@Ejaz, cluster 6 is finished! Without getting too technical, we’re currently using memcache on some of the architecture but the lack of security and authentication features stopped us from making it a user-facing tool.
June 23rd, 2011 at 12:38 pm
I host several websites on my gs account and never had an issue with Media Temple. Thank you for the update and keep up the good work.
July 4th, 2011 at 4:04 pm
Cool!
July 4th, 2011 at 4:04 pm
Very interesting article!
Thanks you very much!)
August 16th, 2011 at 10:19 am
This is quite an old article. I dont know if i will get a response but i will post anyway.
1. I just rethink and reread everything. So is (cs) still going to be a Low Level services where all users are expected to upgrade to (dv) ? And it wont be scalable like Storm on Demand / Rackspace Cloud?
2. Memcache / Percona DB Server are these coming?
3. RAM increase? I guess this has to do with memcache?
4. I heard PHP APC has already been turned on in Grid, am i correct?
August 16th, 2011 at 10:36 am
@Ed, Percona DB is being used in the background on the (gs), and memcached is running as well. We’re managing them on an infrastructure level to increase performance across the board. Email me directly or, better yet, head into the forums to discuss the rest!