Earlier today, I sat in on a keynote presentation at Salesforce.com’s analyst event in New York City. CEO Marc Benioff and other Salesforce execs went over the earlier news that companies can now track Twitter conversations inside Salesforce. Naturally, I Twittered my notes (reproduced below). Salesforce is basically implementing Track (the ability to search and monitor conversations by keyword and topic) inside Salesforce.com in a way that hopefully Twitter will make possible for all of its users.
But the data point I found most interesting had nothing to do with Twitter. Salesforce talked about its own back-end infrastructure and revealed that all of Salesforce.com runs on only about 1,000 servers. And that is mirrored, so it is really only 500. Think about that for a minute. Salesforce has more than 55,000 enterprise customers, 1.5 million individual subscribers, 30 million lines of third-party code, and hundreds of terabytes of data all running on 1,000 machines. Amazon’s Web Services, in comparison, runs on about 100,000 machines I am told by someone with knowledge of Amazon’s server infrastructure.
The comparison is not entirely fair because some of Amazon’s Web services, such as its EC2 compute cloud, are not shared among customers. (In other words, when a developer signs up for it, he gets a dedicated machine or portion of a machine running his compute “instance”). But still, that is roughly a 100 to 1 efficiency advantage that Salesforce has over Amazon’s cloud. It gets this by running a proprietary codebase, proprietary database, and proprietary “multi-tenant optimizer” that slices and dices the data in a very efficient way.
All of Salesforce relies on data stored in only ten databases that run on about 50 servers. It holds several patents on ways to index the tens of millions of rows of raw data. But it’s secret weapon is that “optimzer” which queries the databases and makes sense of all the data. This is all highly proprietary stuff. Benioff pooh-poohed open-source efforts that are less efficient: “We have real-time query optimization. We don’t use some out of the box open source query optimization. Those things don’t work.” Ouch.
Below are my Twitter notes from the event in chronological order (bold added for emphasis):
# Listening to Marc Benioff at press/analyst event. Says Salesforce has 30M lines of code written by others via APPExchange.
# Marc Benioff claims writing an app on the cloud is 5X faster and 5X cheaper than creating conventional enterprise apps. Good talking point.
# Salesforce stole Genius features from Apple, shows related deals. Benioff loves to borrow from consumer apps and make them enterprise.
# Wow, customer service support is now an $8B business, up from $5B in 2004. But customer satisfaction is flat. All going to the cloud.
# At Salesforce event, they are finally talking about Twitter. They claim 8M Twitter users. Wonder where they get that from.
# Salesforce CRM for Twitter is basically Track. Lets you monitor, search, track topics inside salesforce.
# Salesforce Twitter Track searches both original Tweets and replies, even if the keyword is not in the reply. Twitter, please take note.
# Salesforce Twitter feature imports the entire conversation when it gets a search hit.
# Salesforce Twitter integration is two-way. Cos. can monitor Twitter conversations, and then reply back, and the reply appears on Twitter.
# Benioff on need for new development environments for cloud apps: “These are not things you buy out of the box from Frys!”
# Salesforce hosts 13M customizations to its apps on its database.
# Salesforce has over 30M lines of 3rd party code. How does that compare to how many lines of code are in an Oracle or SAP app? Anyone know?
# Starbucks CTO saying they have gotten 65K ideas from customers in a year, MyStarBucksidea. Only 25 ideas implemented (like anti-spill stick)
# Starbucks CTO Bruzzo: tried to raise 1M hours of volunteer service in January both from stores and online. Built app in 21 days.now 1.3M hrs
# Benioff: “We have real-time query optimization. We don’t use some out of the box open source query optimization. Those things don’t work.”
# Salesforce enterprise customers can open up tunnels and share data with each other. Cool. It’s EDI for the masses.
# Salesforce manages hundreds of terabytes. Salesforce runs less than 1,000 machines. Salesforce running on about 10 databases worldwide.
# Each Salesforce database supported by about 50 servers, 2 mirrors, one codebase.
# All of the data of all the billions of rows in salesforce fits into only about 20 tables. Patents on indexing pivot tables.
# Salesforce takes the raw data and compresses. Creates a “multi-tenant index”: “shared massive structures with tens of billions of rows.”
# Salesforce’s secret sauce: It queries its databases with “The Multi-Tenant Optimizer.”
# Salesforce database going from tens of millions of rows to billions of rows per tenant/customer.
(Photo by JohnSeb).
This is an awesome post and quite a data point. I wish we could get some more insight into how the servers are configured to cover such a load. It is pretty incredible how they index the large amount of data they have. I wonder what kind of boxes they are using for this. Very cool, thx for the post!
dell poweredge servers?
no, that was just a generic server (CC) image I grabbed.
Maybe 1000 servers working at one layer of the stack … but I am sure they have more than 1000 database servers.
Eric, I built a 10 server cluster for about $7,000
http://picasawe...reRequirements#
That’s including a layer-4 router with twice the power of the latest Nortel Alteon load balancers
in 2 months we’ll have another one in NYC.
These 20 dual Xeons will support 1 million CDN clients on 2 gigabit small factor fiber optic connections.
Computing power is not only measured by straight math, you have to count that not every client is in the same time zone and they won’t burst at the same time.
They could have done it with 50 servers on the same load. Plus you have virtualization. Something tells me they haven’t figured out how to load balance SaaS based on packet routing.
Here at Techcrunch you are mistakenly running Varnish reverse proxy instead of routing packets to low load servers based on their port number and 8 byte headers.
Who ever engineered the Techcrunch cluster is probably of the same caliber as the Salesforce guys that need 1000 servers.
Pay low and OPTIMIZE!!!
But hey, I’m a pro. It’s not really fair is it?
“in 2 months we’ll have another one in NYC.”
I mean another 10 x 2U dual xeon cluster with a layer-4 balancer in front, not another single server.
“8 byte headers.”
I of course meant 8 bits, an 8 byte TCP/IP header would be crazy.
Chris I am interested in your network building skills but the picasa link you gave doesnt work. How can I see some of your work?
Thankfully TC found at least one thing interesting that had nothing to do with Twitter: *the data point I found most interesting had nothing to do with Twitter.*
L_M_AAAAAA_O
Except those of us who read articles with the hope of finding thoughtful writing rather than a real time spew are now finding the supposed articles are regurgitated Twitter.
One of the most important features of the written word is that production generally requires greater effort than consumption. An author may spend hours composing an exposition that takes only moments to read. That is how knowledge is created.
I don’t understand what you are getting at with “only 1000 servers”. What technical experience are you using to make a comparison, on how much is too much or too little when it comes to a server usage. Those 1000 machines are handling all of the front end work anyhow. The terabytes of data and billions of rows are being handled by those database servers.
I’ve certainly seen obscenely large systems being run on only a handful of machines before.
What they didn’t mention was the large SANs they must have to deal with that much data.
this ^
Plus the comparison to amazon services was lame. Shows a lack of tech knowledge.
Erick, I’m sure you meant “About” instead of “1bout”
in the 2rd sent. – 2rd paragraph. It currently reads, “Salesforce.com runs on only 1bout 1,000 servers”
I am sure you meant 2nd instead of 2rd as well. Typos are not fun…
You’re right.
Thanks
While it certainly makes for a neat soundbyte to note that Salesforce uses 1,000 servers while Amazon Web Services uses 100,000 servers, this comparison essentially useless from a technical perspective without a substantial amount of context.
As other comments above have noted, the number of servers thrown at a problem may have little or nothing to do with efficiency. If those servers are high-end, expensive boxes with premium parts or low-end, “throw-away” hardware we are talking about two completely different things.
More importantly, what Amazon and Salesforce are doing are not comparable at all. Yes, they are both SaaS offerings. That is basically where the similarity ends. AWS is solving a problem that is orders of magnitude more complex and resource intensive.
TechCrunch provides a terrific source of news and interesting reports, but should probably hold off from drawing technical conclusions…
I am not drawing any technical conclusions. I know it is not an apples to apples comparison and say as much. It is an interesting data point, that is all.
No worries – I’m not trying to knock you, just to provide more context around the report. Both data points (the hardware deployments for Salesforce and AWS) are very interesting by themselves since they hint at architecture decisions. I just wanted to make the point that “efficiency” is really not a metric that can be inferred. Please keep up the good work in tracking down intriguing facts and stories.
It is also worth noting that AWS said this time last year that they had 400,000 users which has no doubt substantially increased. Also there is no accounting for the AWS users who decide spend money on more servers instead of optimizing code. Remember that SF is a more unified code base with external systems that make API calls into it. These external services are, in many cases, fairly slow and inefficient compared with the core service. AWS is a bunch of different customers who run their own apps in the cloud most of which don’t interface with each other.
Very interesting comments on the architecture of SF though.
eric, i’ll try to summarize what you’re critics are tying to say in as few words as possible:
YOU ARE AN IDIOT
“what you are critics…”? Yes, you are an idiot.
None of this matters, you can’t benchmark either system for throughput.
These statements are purely marketing.
You can put a grilled cheese in a 4U Sun Netra case and call it a server.
What we’re going to do is let potential customers download 1MB of packet data and measure their speed.
CDNs don’t even allow people to do that. We will.
Europeans will not get great speeds until we have a region there, BUT
at least we’ll let people conduct metrics.
Why doesn’t Amazon, Akamai or Salesforce allow people to conduct speed tests? Any request should be forwarded to the nearest region by DNS, then the layer4 routing kicks in for HTTP/service.
That’s how it *should* work. That’s not how it more often than not works.
So we will let people measure or fiber optic lines. So should they. I bet our US national rig beats them. I’m counting on it.
What? Pipe down.
What? shut it sol
Can every post please start with “Cloud”..please, please, please!
This is hilarious… what a meaningless article. As many times as I’ve seen Erick post an article where he misconstrues cloud computing, though, I’m not real surprised.
your comment would be better if added to mine (above)…
erik has huge problems relating or comparing things to form an opinion that others might find useful…
So this is why their performance sucks…
I hope they add a few more servers soon (maybe borrow some from AWS), because the salesforce.com app is really slow and that affects our salespeople’s productivity.
interesting numbers
Salesforce has a data-intensive application, not a CPU-intensive one. Agree w/ Jason, the real computing challenge is their SAN architecture. Understanding how many of those 1000 machines are used for caching and query optimization would be telling…
Wow, that is surprising. Any talk what makes them so efficient compared to others?
Would be surprising if it was actually true.
The way they index and query the data is geared towards a multi-tenant architecture, that’s seems to be the key.
all saas providers, by definition, adopt a multi-tenant architecture.
indexing and querying data efficiently is key to any service, not just salesforce…and their key isn’t any prettier(or efficient) than others…not based on this data anyway.
marc benioff = salesforce = entitled to free speech.
erick schonfeld = techcrunch = obligated to report free speech as is, or make credible/informed observations about said free speech.
not all saas providers are multi-tenant, not even all AWS services. EC2 runs virtualized instances of a server for each customer. It is not one codebase either because AWS lets customers run their own code. It is a different business, but one that requires a lot more servers.
@Erick
You’re hopelessly out of your element. Please don’t embarrass yourself any further by commenting.
Your comparison of AWS and Salesforce is silly – and I’m trying to be as nice as I can be.
It is like me saying Larry Ellison’s house is bigger than Bill Gates’ laptop. Meaningless. Completely.
Erick,
Maybe I am sleepy, but I think you just blurred the line between a Software as a Service provider and an Infrastructure as a Service provider.
Nonetheless, it is very easy to get lost in “the cloud”, given that it’s the new word these days. So it’s totally worthwhile doing a memory refresher on what the “cloud” is comprised of, so we don’t end up getting into redundant apples vs. oranges conversations.
http://en.wikip...ting#Components
My sentiments EXACTLY. You’re so right on. A company running their own servers with their own codebase to provide *software* as a service is certainly nowhere near comparable to a company providing servers as a service. Amazon isn’t running an application that requires hardware, they’re simply providing hardware for others to use. Chances are salesforce.com would need to use just as many machines if they were backed by AWS.
Does seem like an apple-orange comparison with little relevancy. Salesforce is B2B while Amazon is B2C and serving apps, most of which are B2C.
Salesforce has 1.5 million active users, less than Techcrunch. Does Techcrunch have 1,000 servers?
Does not take away from Salesforce’s accomplishments. They do make money.
Good report on their data center, though.
Amazon EC2 is a completely different service. No surprise they use more computing power. One would think a writer for a technology blog would know that.
“The comparison is not entirely fair because some of Amazon’s Web services, such as its EC2 compute cloud, are not shared among customers.”
Do you know how to read?
Then why even mention Amazon in the article if they are totally unrelated? Do you know how to write?
They are related, just not on an apples-to-apples basis. There are very few cloud computing platforms out there operating at scale. Salesforce and Amazon are two of them. It is certainly interesting that one of them runs on 1,000 servers and the other runs on 100,000 servers. To my knowledge, neither of those numbers have been reported anywhere. In my book, that is worth a mention.
LOL
Erick,
Don’t get defensive..I say/write/do stupid things everyday myself. It’s all good in the hood.
But:
If you genuinely believe that the comparison isn’t fair ONLY because of some AWS resources not being shared, then I recommend you do your reading once again.
The comparison is not only unfair, it is redundant, it has absolutely no meaning/relevance, and is extremely uneducated- and that last bit bothers me a lot more than the others.
But I’m willing to listen: what, specifically, did you find so interesting? Take out the Amazon component…in of itself, is there something specific about x customers/Y TB data running on N servers that you found enthralling?
Because a lot of us deal with similar situations everyday and believe you me, it is not exciting at all…and best left to the confines of an excel spreadsheet with good estimation formulae.
if you scroll down to a comment by one of your readers(Kaiyzen), you will see what i mean with the whole excel sheet thing:
“So if you take the 1.5 million claimed users at an average of 10% you have 150K users hitting the system…, again rough averages.
This means that with 500 servers (the mirrored aspect eric pointed out) they are supporting 300 users per box.
This number is neither good nor bad. “
You don’t ‘mirror’ servers in the same way you do drives or data.
The data most likely sits on a NAS or 50 which probably doesn’t count into that 1000.
You won’t have 500 servers sitting idle waiting for one of the other 500 active servers to go away.
You’ll have 1000 servers all active, and when one goes away it just doesn’t get requests anymore. The user never notices because they aren’t connected to any one server, they connect to a load balancer, which then connects them to a specific, operational server.
Since the servers themselves don’t actually store any data (thats what the NAS is for), when one goes down it doesn’t matter other than now more requests go to other servers.
So you take 150,000/spread across 1000 servers and you get 150. For the service that saleforce provides, 150 users per server is rather sad. Take into account that its going to be more like 50/server since we’re talking about 1.5m gobal customers across several time zone, only about a third of them are going to be active during any 8 hour period. Keep going further with the fact that in most SaaS offering like this, you end up with a pretty high number of shelfware seats, so there, you’re probably already down another 25%
So yes they have a lot of data, and also yes, they have what would at first glance appear to be a shitty setup. Google does far FAR more per 1000 servers than Saleforce is likely to ever see total, and they have the entire Internet to search through for text based results.
Sorry, but this just shows how unimpressive something can be technically, but still amazingly successful. Twitter is the same way. Welcome to Web 2.0. Its like MS OS bloat, except it sounds impressive and it comes with free upgrades!
that’s what we do too but it actually created bigger headaches in terms of scalability then if data would be separated into independent silos.
let’s not get personal guys….”cloud computing” is still fairly new, and with the way salesforce screams about the cloud, it almost had someone like ME fooled.
it’s up to the rest of us who work with the cloud to politely beg to differ/make rectifications as needed.
let’s not go for the jugular right away, ja?
but Erick: a little homework wouldn’t hurt either, especially when you’re listening to “information” from the ceo of SALESforce.
we all know the reputation that SALESmen bring with them, right?
Preetam, we know you are trying to demonstrate your half-baked knowledge here… got it got it. Now get lost.
I love TechCrunch. They even let my company (BrowserMob) present at their recent cloud computing roundtable.
BUT…
The folks at TechCrunch need to get educated about the “cloud”. It was evident by the moderators at the roundtable, and it’s evident now. As the other commentators have pointed out, website != platform != infrastructure. Comparing SF.com or Ning or Facebook to Amazon EC2 is absurd.
Both groups are definitely changing the game, but please understand that they are doing it in VERY different ways. Doing so would make your reporting that much more valuable to many of us
That said, keep up the otherwise great work!
Interesting datapoint, but really, BFD. AWS vs salesforce is like comparing apples to, well, some other vegetable matter that may not even be a fruit.
1.5 million individual subs makes salesforce something like 1/100th the scale of Facebook. Probably no one knows exactly how many individual users AWS’s supports, but it starts with all of amazons retail customers, doesn’t it, and all their 3rd party sellers, and all their affiliates, and all their AWS customers, and all their customers.
We also don’t know what kind of servers salesforce runs on, but I’ve heard that their database layer runs on seriously high end iron, not the commodity clusters most web sites are using.
As for efficiency, I wouldn’t be surprised if salesforce is more efficient, since i’m sure they have a lot more visibility into the details of the apps running on their infrastructure than AWS does.
Maybe it’s time to add additional servers. As a savvy SalesForce user it’s very very slow sometimes (or even most of the time).
I agree. It’s awfully slow, a lot of the time. I wondered why. Now I know.
I third this! Beef it up SF!
Not really an apples to apples comparison is it? Salesforce is an application platform. EC2 is an infrastructure platform. Sliced and diced infrastructure platforms will always wasted cycles because of the service levels that come with the provisioning. Salesforce is an application and so it can be architected to maximize efficient use of hardware without the customer’s knowledge. It will be a fair comparison once the xen community and AWS figure out how to distribute a VM’s state and execution based on load.
to your point:
it’s relatively straightforward doing load-distribution when the impacted component is either storage(s3) or transactional(sqs/sdb) in nature.
but if you’re talking about computational resources(EC2), then optimization is a huge challenge.
so again- it’s important to keep in mind which specific aspect of the “cloud” is being used as a basis for comparison.
Preetam, your half-baked knowledge has been acknowledged. Now go eat some apples. dont touch the oranges kid.
Aaam kha.
Seb or Santra chor de Gadhe.
Seriously, TC needs more writers.
Umm, it would be great if the reporter here actually asked someone what it’s like to develop on Force.com. Because it’s not pretty.
You need to wrestle with “governors”, like the kind put on ATVs so kids don’t go too fast. You’re limited to a small number of DML statements per method. You’re limited to how many rows you can query at once. You’re limited to how many API calls you can make per hour (it’s not very many). Building a Force.com app is an exercise in hackery and workarounds. It’s like you’re programming for a shared mainframe in the 60’s.
In other words, running on only 1,000 servers isn’t a feature, it’s a horrendous bug.
Haha. I think that brings us down from the cloud to ground reality…always a good pill.
Thanks, Luigi…this is good to know.
Preetam baby, Now go eat some apples. dont touch the oranges kid.
yes, good point. that is the downside, you are limited in what you can do.
Excellent points here, I was wondering if running on 1000 servers had a downside also. Sounds like they cut down on features/user experience in order to achieve this. Thx for the comment, definitely gave a good perspective.
Just saw your post after submitting my comment. You’re right, it isn’t pretty. It downright hurts, agonizingly so.
I second that. I’ve written Apex in Salesforce and Java outside of Salesforce to access the API, and ran into these limits, and they suck. The worst thing is that the people I write for pay a lot of money for their Salesforce accounts, and then they have to pay me extra to work around the stupidity. Why doesn’t Salesforce just add CPU time to the bill and make life easier for everyone?
Amazon is much more efficient than Salesforce in the “developer time required to make something work” department.
A thousand servers.. about right for the purpose..
as far as number of Servers they are running Blade Servers not 1U. Big difference….
Nonetheless, Salesforce has done a nice job and they have a great business.
Not to continue to harp on the data points laid out in the article, but a few things to point out regarding a web based enterprise application.
Typically for a CRM app you are looking at around 10% concurrency for usage. That number can go as high as 30/40% for call center setups and as low as 1% for the SMB market and general usage.
So if you take the 1.5 million claimed users at an average of 10% you have 150K users hitting the system…, again rough averages.
This means that with 500 servers (the mirrored aspect eric pointed out) they are supporting 300 users per box.
This number is neither good nor bad. Scale things up to a consumer site that is heavily trafficed and you have a since why Amazon, Yahoo!, etc have tens of thousands to hundreds of thousands of servers.
Multi-tenant doesnt really give much benefits when it comes to serving web requests so that is pretty much not relevant.
Sanz sensationalism in this article it would be nice to know what is going on with their caching layer. My guess is a lot of caching on the db side of things, and a fair amount on the web site.
Good point.
And I’d also like to remind you, Erick, that you titled this post “The Efficient Cloud….”.
Sheesh! You’re comparing apples and oranges here. Learn, or research, a little bit about computer science and disparate resource utilization profiles by different kinds of application workflows BEFORE your spew forth more clueless dross labeled as “journalism”. Next.
Work loads are different and it’s hard to call one “cloud” efficient without comparing its work load to others.
For instance, Mozilla supports ~250m users worldwide on less hardware:
mysql> use inventory_mozilla;
mysql> select count(*) from systems; +———-+
| count(*) | +———-+ | 845 | +———-+ 1 row in set (0.00 sec)
But that’s somewhat meaningless as a metric of efficiency.
The Amazon comparison is nonsensical. Amazon in the analogy would be their server farm (albeit virtual), rather than Salesforce itself.
Salesforce provides a platform to a wide (but still limited a targeted) audience that can include structured 3rd-party “applications.”
I mean no offense here, but this is basically not a “wow” moment.
@Nathan – you nailed it. This is the stupidest comparison EVER.
One of United Airlines Jumbo 747 can cart 500 passengers across the US in just 4 hours. you would need 400 cars from enterprise or hertz by comparison to do the same and it would take 4 full days!
So what – EC2 is a true Cloud Platform – able to host arbitrary workloads on demand. Salesforce is a single purpose Application platform able to host only a specific set and subset of capabilities.
Tu Gadha hai.
From your notes…
# Salesforce manages hundreds of terabytes. Salesforce runs less than 1,000 machines. Salesforce running on about 10 databases worldwide.
# Each Salesforce database supported by about 50 servers, 2 mirrors, one codebase.
If you use the 10 databases * 50 servers * 2 mirrors you get 1000 servers just for the dbs…I won’t even go into the Amazon comparison…more sensationalistic nontechnical techcrunch baloney
Unreal. I love comparisons like this.
“Amazon’s Web Services, in comparison, runs on about 100,000 machines I am told by someone with knowledge of Amazon’s server infrastructure.”
Erick – whoever told you that has *no* knowledge of Amazon’s server infrastructure. That number is so far from the truth as to be laughable. You are doing the most reprehensible thing a “journalist” (and I cringe when I type that word) can do — you are making shit up.
TechCrunch – you really need to ditch Erick’s ridiculous articles. They are the worst case of mumbo jumbo nonsense that are not based on fact, but conjecture (and in the case of the data provided in this article – downright lies). They sully what is a fairly reasonable site.
The 1,000 servers could also be multi-processor cores running various configurations of VMs. So the actual number of instances could be much higher. This is not global class. Google had over 1 million servers 18 months ago according to Gartner. Amazon, Google and Microsoft have 5 million servers or more today.
One of the reasons why Google has so many servers is that the entire index is held in memory. Hence the impressive speed of the search results.
Of course the servers are multi-processor cores. Who would deploy anything else these days? But Erick was not writing about “Instances” – he wrote about “machines”: “Amazon’s Web Services, in comparison, runs on about 100,000 machines” — which is flat out incorrect.
A lot of unfair & harsh comments here.
Anyway, a Facebook/Salesforce hardware vs. revenue comparison would have been more like-for-like because Salesforce makes up to $200/user/month, wheveras Facebook would only dream of $1/user/month.
Not really to educated on the topic as it is new to the block. Regardless of the “errors” in the article and the weird comparisons, I would now consider myself an expert on the topic after reading all the comments.
While Erick may not be the best suited person to write this article, he has the power to bring us all together which creates the most value. I dont think he expected this much criticism when he wrote the post, but we live and learn because it cant be that easy to write to thousands of people without someone getting pissed off. Keep it up!
Agree with James above, revenue / server is kind of an interesting statistic. Of course, you’d have to know the class of hardware involved, as others have stated, then you could do something like a revenue / $$ of capital costs calcuation, but it would surprise me if you couldn’t figure something like that out from quarterly filings.
At any rate, SFDC has historically been very forthcoming about their infrastructure. There are presentations from Claus Moldt on the developer site explaining the different tiers and going into some detail about what class hardware runs at each. Also there’s always http://trust.salesforce.com if you’re interested in seeing how many transactions are being handled, and at what rate. More SaaS companies should be so open about their platform.
There is lot of high value data inside SFDC but it is just not a high transaction site – reflecting the nature of how CRM applications are used in practice.
A typical day has 190,165,910 transactions if I read that right on the status page. Over a 12 hour busy day that is 4400 transactions per second across the 500 or 1000 machine cluster. 4-8 transactions per second per server… ? Yawn.
They posted the video they showed;
http://www.yout...K_ms&fmt=22
It’s not very technical but at least I can send it to my boss to explain the basic concept.
John
Salesforce actually published quite a few details on how the architecture works a few months back:
http://wiki.dev...nt_Architecture
With force.com, Salesforce is much more than just CRM, and I would imagine that as customers adopt the ability to run arbitrary code, they will add more servers.
I don’t think the feature is GA yet, but when it is, the Force.com sites feature could also require them to buy more servers:
Check out some of the sites here:
http://blog.sfo...-challenge.html
I think the number of computers isn’t a fair comparison, the amount of computing that the system can do and how efficiently the resources are used is a better comparison.
An interesting discussion would be if Salesforce ran on EC2 or something like it, how many instances would it require and could that even be done, and could salesforce lower their hosting costs by doing this, and whether Amazon or someone else has security standards high enough to meet the requirements of the larger customers that Salesforce serves.
While I can see this as a cost advantage to SFDC, I am struggling to see how this translates into a cost advantage for people using force.com.
The barrier I see for SFDC is that one has to develop using force.com. AppExchange is only useful for people who want to integrate to SFDC and other companies using AppExchange. But it isn’t – in my opinion anyway – a development platform. So the real measure of SFDC becoming a big player in PaaS is the adoption of force.com as a development environment. Has anyone seen numbers on this? Until they can get people using force.com their cost advantage isn’t leveragable and the early adopters have to invest in new development capabilities and training. As an example, I doubt that we could get our Java UI to run in force.com, so our adoption costs would be as high as, say, moving to a .NET UI.
It is tough to see quite who the winners will be in the long-term, but in the short-term I think it is more likely going to be the IaaS plays, such as Amazon. They allow companies to leverage their existing development skills and existing code in a more cost effective manner
Could you please point us to some benchmarks? Proprietary optimization algorithms sounds like a bunch of crap to me. They’re probably running oracle farm with a bunch of dells, which to this date is the fastest cluster configuration out there.
Nice article, good data points.
See also a recent attempt to estimate infrastructure implications of tweeting.
http://yescloud...ns-of-tweeting/
If the underlying assumptions there are in the ball park, the server/storage need appears quite manageable, with a one to two orders of magnitude buffer…
PG.
Prashant, I’m not sure what type of curry you are on, but there is hardly any data provided to make meaningful conclusions about whether salesforce infrastructure really scales.
“Marc Benioff claims writing an app on the cloud is 5X faster and 5X cheaper than creating conventional enterprise apps. Good talking point.”
Good talking point, but far from reality. I just spend three months building an app integrated with the Salesforce platform. Our app interacts with PayPal and an external website.
SOQL (Salesforce Object Query Language) leaves a lot to be desired (top of my list would be SELECT *). SOAP connection times to the app are on the order of 1-5 seconds, making it somewhat untenable for transactional usage.
Working within the Salesforce app via browser is painful, 1-5 second page load times there. Forget your password? That’s four emails you have to delete Need to alter an object attribute? 20 seconds worth of clicking. Doing so with vim, emacs, or your favorite GUI editor the traditional way is undoubtably faster. Debugger? Sorry.
This platform is decently usable for a class of application developers, but it may cause your programmers to run away and hide. Writing code to interface with the platform is painfully slow and agonizing. I say this from experience.
It could be that I am an outlier in that I write code for a living. But if I compare the way I traditionally write applications (using an interpreted language and an application framework), the difference with the Salesforce cloud is that it is 5 times slower and 5 times more expensive, not the other way around.
Second you! Painful Painful and Painful!!
Just finished a project for AppExchange, it was painful to develop, and even more painful to do all the paper stuff.
The whole idea of Salesforce is good but the implementation is not pretty. I wonder how many live customers are actually using features beyond what comes with a standard Sales account.
@Luigi above, excellent point!!
Reposting -
“Umm, it would be great if the reporter here actually asked someone what it’s like to develop on Force.com. Because it’s not pretty.
You need to wrestle with “governors”, like the kind put on ATVs so kids don’t go too fast. You’re limited to a small number of DML statements per method. You’re limited to how many rows you can query at once. You’re limited to how many API calls you can make per hour (it’s not very many). Building a Force.com app is an exercise in hackery and workarounds. It’s like you’re programming for a shared mainframe in the 60’s.
In other words, running on only 1,000 servers isn’t a feature, it’s a horrendous bug.”
I met few people who worked at Salesforce and let me tell you they’re not the brightest bunch.
Does Amazon have 100000 servers !!! I believe it is far lesser
Do you have stats to prove otherwise?
so, yes, many people have already commented about the Amazon – Salesforce comparison. But I still feel the need to add to that: you need to remove the sentence:
“But still, that is roughly a 100 to 1 efficiency advantage that Salesforce has over Amazon’s cloud.”
To compute “efficiency” you need some measure for what these services actually do. Even leaving aside the complexity of the offered services and whether they are shared or not: you only have measures of scale for salesforce, no number of customers, data managed etc. for amazon. You just have no basis for any statement on efficiency; for all we know amazon could be much more efficient than salesforce. From your (otherwise interesting) numbers there is no way to establish relative efficiency between these two.
Stop it, you’re bringing logic into this conversation now. His head will explode.
All you super smart guys, try writing an article of your own. Dont just BS and talk big about Apples n Oranges.
@ronny: you’re sending a lot of love my way…do i know you, darling?
No Jerk. But i do know you. Who sits all day on websites trying to pull down others and show-casing your half-baked knowledge.
It is very clear that the writer of this article just made a casual comparison, and nowhere implying that he was comparing apples.
Now, as i said – go eat your rotten apples and smelly oranges.
Guys – not sure what presentation you are referring to but I’ve been to salesforce.com and I know for a fact after meeting with Parker Harris and their VP of Dev they have WAY more than 1000 servers. Maybe you’re just counting web servers.
He really has no idea what he is counting. He’s a business writer attempting to be technical.