Ok, so the thing is, we have multiple NodeJS web servers which need to be online all the time. But they'll not be recieving many requests, approx. 100-200 requests a day. The tasks aren't CPU intensive either. We are provisioning EC2 instances for it. So, the question is, can we run multiple nodejs processes on a single core? If not, is it possible to run more low intensity NodeJS processes than number of cores present? What are the pros and cons? Any benchmarks available?
Yes, it is possible. The OS (or VM on top of the OS) will simply share the single CPU among the processes allocated to it. If, as you say, you don't have a lot of requests and those requests aren't very CPU hungry, then everything should work just fine and you probably won't even notice that you're sharing a single CPU among a couple server processes. The OS/VM will time slice among the processes using that CPU, but most of the time you won't even have more than one process asking to use the CPU anyway.
Pros/Cons - Really only that performance might momentarily slow down if both servers get CPU-busy at the same time.
Benchmarks - This is highly dependent upon how much CPU your servers are using and when they try to use it. With the small number of requests you're talking about and the fact that they aren't CPU intensive, it's unlikely a user of either server would even notice. Your CPU is going to be idle most of the time.
If you happen to run a request for each server at the exact same moment and that request would normally take 500ms to complete and most of that was not even CPU time, then perhaps each of these two requests might then take 750ms instead (slightly overlapping CPU time that must be shared). But, most of the time, you're not even going to encounter a request from each of your two servers running at the same time because there are so few requests anyway.
I am designing a web site in which users solve puzzles as quickly as they can. JavaScript is used to time each puzzle, and the number of milliseconds is sent to the server via AJAX when the puzzle is completed. How can I ensure that the time received by the server was not forged by the user?
I don't think a session-based authenticity token (the kind used for forms in Rails) is sufficient because I need to authenticate the source of a value, not just the legitimacy of the request.
Is there a way to cryptographically sign the request? I can't think of anything that couldn't be duplicated by a hacker. Is any JavaScript, by its exposed, client-side nature, subject to tampering? Am I going to have to use something that gets compiled, like Flash? (Yikes.) Or is there some way to hide a secret key? Or something else I haven't thought of?
Update: To clarify, I don't want to penalize people with slow network connections (and network speed should be considered inconsistent), so the timing needs to be 100% client-side (the timer starts only when we know the user can see the puzzle). Also, there is money involved so no amount of "trusting the user" is acceptable.
You can't guarantee the security of the timings cryptographically, because the client's browser can't do secure computation. Any means for encrypting to/from the server could be bypassed by adjusting the actual timings.
And timing on the server doesn't work, either - if you don't take account of latency in the round-trip-time, users with lower latency connections will have an advantage; if you do, users could thwart the compensation phase by adding extra latency there and then removing it later.
You can, of course make it difficult for the users to modify this, but security by obscurity is an unsustainable policy anyway.
So it comes down to either trusting your users somewhat (a reasonable assumption, most of the time) and designing the game so it's not trivial to circumvent the timings.
This approach obviously makes assumptions and is not invincible. All calculations are done on the client, and the server does some background checks to find out if the request could have been forged. Like any other client-based approach, this is not deterministic but makes it very hard for a lying client.
The main assumption is that long-lived HTTP connections are much faster for transmitting data, even negligible in some cases depending on the application context. It is used in most online trading systems as stock prices can change multiple times within a second, and this is the fastest way to transmit current price to users. You can read up more about HTTP Streaming or Comet here.
Start by creating a full-duplex ajax connection between the client and server. The server has a dedicated line to talk to the client, and the client can obviously talk to the server. The server sends the puzzle, and other messages to the client on this dedicated line. The client is supposed to confirm the receipt of each message to the server along with its local timestamp.
On the server generate random tokens (could be just distinct integers) after the puzzle has been sent, record the time when each token was generated, and pass it over to the client. The client sees the message, and is supposed to immediately relay this token back along with it's local time of receipt. To make it unpredictable for the client, generate these server tokens at random intervals, say between 1 and n ms.
There would be three types of messages that the client sends to the server:
PUZZLE_RECEIVED
TOKEN_RECEIVED
PUZZLE_COMPLETED
And two types of messages that the server sends to the client:
PUZZLE_SENT
TOKEN_SENT
There could be a lot of time variation in the messages send from the client to the server, but much lesser in the other direction (and that's a very fair assumption, hey - we have to start somewhere).
Now when the server receives a receipt to a message it sent, record the client time contained in that message. Since the token was also relayed back in this message, we can match it with the corresponding token on the server. At the end of the puzzle, the client sends a PUZZLE_COMPLETED message with local time to the server. The time to complete the puzzle would be:
PUZZLE_COMPLETED.time - PUZZLE_RECEIVED.time
Then double check by calculating the time difference in each message's sent vs received times.
PUZZLE_RECEIVED.time - PUZZLE_SENT.time
TOKEN_RECEIVED.time - TOKEN_SENT.time
A high variance in these times implies that the response could have been forged. Besides simple variance, there is lots of statistical analysis you can do on this data to look for odd patterns.
Even a compiled application could be forged. If the user changes their system clock halfway through timing, your application will report an incorrect time to the server. The only way to get an accurate upper-bound on the time it takes them is to start timing on the server when the puzzle is given to them, and to stop timing when they supply the answer.
As others have pointed out you can minimise the effect that slow connections have by making the load of the puzzle as small as possible. Load the entire page and "game engine" first, and then use an asynchronous request to load the puzzle itself (which should be a small amount of data) to level the playing field as much as possible.
Unfortunately you can't do latency compensation as this would be open to tampering. However, on a connection that's not being used for anything else, the latency for a request like this would be greatly overshadowed by the time it takes a human to solve a puzzle, I don't think it will be a big deal.
(Reasoning: 200ms is considered very bad lag, and that's the average human reaction time. The shortest possible "puzzle" for a human to complete would be a visual reaction speed test, in which case bad lag would have a 100% markup on their results. So as a timing solution this is 2-OPT. Any puzzle more complex will be impacted less by lag.)
I would also put a banner on the page saying to not use the internet connection for anything else while playing for the best possible speeds, possibly linking to a speed / latency tester.
It is impossible to start and stop the timer at the client-side without fear of manipulation...
Anything you perform at the client can be altered / stopped / bypassed..
encrypting/decrypting at the client is also not safe since they can alter the info before the encryption occurs..
Since it involves money, the users can not be trusted..
The timing has to start at the server, and it has to stop at the server..
Use ajax to start the timer at the server only if the puzzle contents return with the result of the ajax call. do not load the puzzle and then sent an ajax request as this could be hijacked and delayed while they review the puzzle...
..
Depending on the server side implementation you have, you could put the timing functionality on the server side. Record the time that the webpage request was made (you could put that into a database if you liked) and then when the answer is received get the current time and perform some arithmetic to get the duration of the answer. You could store the time in the session object if you liked instead of the database as well although I don't know too much about its integrity in there.
You have to use server-side time here. Here is how I would do it:
Make an AJAX request on document ready to ping the server. When server-side code receives the ping, store the server-side time as a session variable (making sure the variable does not already exist). When they finish the quiz, take the server-side time again and compare it with the session variable to determine their duration. Remove the session variable.
Why this works:
You do not start the timer before they see the quiz
The network delay is factored in, because the timer does not start until the AJAX request comes in (if they have a slow connection, the AJAX request will be slow)
Ping is not spoofable because you make sure the session variable does not exist before storing
EDIT: I wanted to add that you could continue to keep client-side time, and include it in the final post. Then you can compare it with your server-side calculated time. If they are reasonably close, then you can trust the client time.
You asked a bunch of questions in your original question, I'm only going to answer one of them:
Am I going to have to use something that gets compiled, like Flash? (Yikes.)
Yes. Given your criteria: 1) 100% accurate, and 2) No possibility of user interference, you have to use a compiled binary.
Doesn't have to be flash though - I'd suggest a java applet if the thought of Flash makes you say "Yikes".
-- Edit:
This solution is somewhat flawed, as pointed out by ZoFrex below.
-- Old:
Here is a way (but you'll need to do some profiling).
Send down a series of "problems" for the JavaScript to solve, while they are playing the puzzle. Previously, I've sufficiently-sized number N such that it is the result of: prime1 * prime2. This forces the client to factor the number (you can get code to do this in JavaScript) and this will take time (this is where profiling clients comes in, and sending down appropriately-sized primes [obviously, this opens you to degradation-attacks, but nevertheless]).
Then, you just send down say, 500, of these prime-problems (or another type), and let the JavaScript solve them in the background. It will generate a list of solutions, and when you send the completed value, you also send this list. From the total count of answers supplied, you can determine how long they spent on the puzzle.
Cons:
Requires profiling to determine capabilities of various clients (and hence difficulty of problems)
Can be downgrade-attacked
Slightly complicated
JavaScript computation may interrupt general puzzle-solving
Possible to write a bot to get solve problems faster than JS
Pros:
Calculations must be done in order to submit the form
If implemented correctly, will prevent all but non-trivial attacks
Clearly, it's attackable, (all proposed answers are), but I think it's reasonable. At least, it would be fun to work on :)
In the end, though, you need to actually install a client-side system with a bit more security. And do note that Flash certainly is not this; it's trivial to decompile. Infact, there was an IQ test here in Australia once, and it was controlled via a Flash app that was done LIVE on television. Of course, the winner was a computer programmer, I wonder why :P
-- Edit:
OP, Also, I linked it in a comment to this post, but just incase you miss it, you are kind of interested in the Hashcash, which is the aim to show that a client has completed some amount of 'Work'. Even if my implementation isn't suitable, you may find a review of that field fruitful.
It's a tricky problem because it's fundamentally unsolvable, so you need to work around the tradeoffs to do your best. There've been several good points made on the technical side including: (a) don't waste your time thinking compiling to Flash, Windows, Silverlight, JVM, or anything will actually help, (b) first transmit the encrypted real puzzle payload, then as the actual bottleneck transmit the key alone, (c) the latency even on 56k of sending a couple hundred bytes is negligible compared to human reaction time.
One thing I haven't seen mentioned is this:
Use after-the-fact auditing and tracking of users. This is how real casinos work. This is, I am told, a big part of how PayPal made their money. In both cases, rather than doing a lot of before-the-fact security, they keep very close tabs on everything about their players, use a lot of technology (statistics, pattern detection, etc) to flag suspicious players and they investigate. In both casinos and PayPal, you can't get your money right away. You have to cash in your chips or wait for the money to make it out of the PayPal system into your real bank. Your system should work the same way-- they can't actually get the money for at least a few days at minimum (longer if you are unable to set up a sufficiently speedy auditing system), giving you time to potentially impound their winnings. You have a lawyer, right? Also, casinos and PayPal know your real life identity (a must for dealing in money) so they can go after you legally-- or more importantly, deter would-be attackers since they could go after them legally.
Combined with the other tips, this may be sufficient to eliminate cheating entirely.
If you find it is not, make your goal not to totally eliminate cheating but to keep it to an acceptable level. Kind of like having 99.99% uptime. Yes, as one poster said, if even one person can compromise it everyone is screwed, but with a good auditing system the attacker won't be able to consistently cheat. If they can cheat 1 in 1000 times if they're lucky and they find they can't cheat more than once or twice before being caught, it won't be a problem since very few will cheat and any given honest user will have an extremely low chance of being affected by an extremely small amount of cheating. It'll be imperceptible. If you ever have a real cheating occurence that hurts an honest user, go out of your way to make the honest user feel satisfied with the outcome to a degree out of proportion to the value of that single customer. That way everyone will be confident in your security. They know they don't have to worry about anything.
People problems are not always solvable with technology solutions alone. You sometimes need to use people solutions. (The technology can help those people solutions work a lot better though.)
excuse me but why you can't use the time on the server? the time when you recieve the response will be the one which you use to calculate the score.
As several others have pointed out:
You must use server time, because client time is vulnerable to manipulation.
Checking the time on the server will potentially penalize people with slow network connections, or people that are far away.
The answer to the problem is to use a time synchronization protocol between the client and the server similar to the protocol that NTP uses. Working together, the client and the server determine the amount of delay caused by network latency. This is then factored into the times given to each user.
NTP's algorithms are complicated and have been developed over years. But a simple approach is below; I think that the protocol should work, but you may wish to test it.
Have the client measure the round-trip time with two successive HTTP XMLRPC pings. Each ping returns a different nonce. The second ping requires the nonce from the first ping, which assures that they are sequential. The puzzle time starts when the second HTTP ping is sent from the client. The server timestamps each request and assumes that the puzzle is displayed 1/2 way between the receipt of the first and the second request.
When the puzzle is finished the client pings twice again, following the same protocol as before. The server knows when it receives each request and it knows the time delta. Now take half the time delta and subtract that from when the first ping of the second set is received. That can be safely assumed to be the time that the puzzle was completed.
there is a very fast implementation of cryptography in js here
http://crypto.stanford.edu/sjcl/
it allows public / private key encryption all on the client and I think you can adapt it to encrypt the Ajax communication between your server and the client browser
here is a detailed explanation, which you can adapt to your needs
http://crypto.stanford.edu/sjcl/#usage
Just a quick thought: why don't you use an iFrame to include the game and it's javascripts and let them reside on the server you have your server side implementation running. Any ajax request should then be sent by the same IP as your server side IP is which would solve the problem of identifying the source. Of course you have to take further measures but already gained a lot of confidence in your "client" side requests. Remember the windowsLive services login and many more like it are based on javascript and the usage of iFrames and are considered secure enough.
I do not think there is a perfect solution. Here is an alternative that makes it harder for cheater but at the same time an unlucky honest solver may lose out.
Get many samples of roundtrip time measurements from each specific devices/location/other combination apriori for each user based on their other interaction with your site. You will also have these measurements for the entire population. You could also be very subtle get the timestamps for when a particular DNS lookup happened from their ISP's resolver (random hostname and you host the authoritative DNS server for that domain).
Once you have this, perform all measurements on the server side (puzzle returned to user, solution received) and subtract out the network time based on previous observations.
Note that even in other solutions you have server load, client load (slow processor, etc.), etc that affect timing.
Make sure you have XSRF protection on puzzle submission page :)
The way I would do this is that when the server sends the puzzle to the client, the current time is stored in a session. This ensures that the timing starts immediately after the puzzle has been sent. After the puzzle has been completed, and is sent to the server to check if the puzzle was done right, the server again checks the time and does a comparison.
Obviously slow Internet connections can make this time bigger, but there's nothing you can do about it.
Is there a way to keep a HTTP connection alive with JavaScript?
In short, I think the concept of long lived http connections in javascript really revolve around a style of communication called COMET. This can be achieved in several different ways, but essentially involves the client (using XmlHttp powers) requesting data from the server immediately, and the server withholding the response until some event triggers it. Upon receipt of this response, the client immediately makes another request (which will once again hang at the server end until something needs sending). This simulates server push, but is effectively nothing more than a delayed response used in a clever way. In the worst case, there can be fairly high latency (i.e. 2 messages need sending, so the cycle must be twice repeated, with all the costs involved) but generally, if the messaging rate is low, this gives a reasonable impression of real-time push.
Implementing the server-side for this kind of communication is far from trivial, and requires a good deal of asynchronous communications, concurrency issues and the like. It's quite easy to write an implementation that can support a few hundred users each on their own thread, but to scale to the thousands requires a much more considered approach.
I note that the last answer was given in 2009. Oh, how I remember the days. But lots of good things have happened since then; so I'll add this just to let people know what to look for. HTTP 1.0 provided a "keep-alive" request property that meant that the connection should be maintained for further requests. In HTTP 1.1, this became the default. You actually have to opt-out of it if you don't want to reuse the connection (and if you want to be nice about it).
The new standard for "WebSockets" actually gives you a full-duplex persistent connection. WebSockets are supported in all up-to-date versions of popular browsers and you can even use them in MSIE if you install the Google Chrome Framework (which means Google software is actually doing the work). Microsoft says IE supports it in version 10, but I haven't tried it myself. What you need then is something to connect to, like http://highlevellogic.blogspot.se/2011/09/websocket-server-demonstration_26.html
I am working on a simple notification service that will be used to deliver messages to the users surfing a website. The notifications do not have to be sent in real time but it might be a better user experience if they happened more frequently than say every 5 minutes. The data being sent to and from the client is not very large and it is a straight forward database query to retrieve the data.
In reading other conversations on the topic it would appear that an AJAX push can result in higher server loads. Since I can tolerate longer server delays is it worth while to have the server push notifications or to simply poll.
It is not much harder to implement the push scenario and so I thought I would see what the opinion was here.
Thanks for your help.
EDIT:
I have looked into a simple AJAX Push and implemented a simple demo based on this article by Mike Purvis.
The client load is fairly low at around 5k for the initial version and expected to stay that way for quite some time.
Thank you everyone for your responses. I have decided to go with the polling solution but to wrap it all within a utility library so that if they want to change it later it is easier.
I'm surprised noone here has mentioned long-polling. Long polling means keeping an open connection for a longer period (say 30-60 seconds), and once it's closed, re-opening it again, and simply having the socket/connection listen for responses. This results in less connections (but longer ones), and means that responses are almost immediate (some may have to wait for a new polling connection). I'd like to add that in combination with technologies like NodeJS, this results in a very efficient, and resource-light solution, that is 100% browser compatible across all major browsers and versions, and does not require any additional tech like Comet or Flash.
I realize this is an old question, but thought it might still be useful to provide this information :)
Definitely use push its much cooler. If you just want simple notifications I would use something like StreamHub Push Server to do the heavy-lifting for you. Developing your own Ajax Push functionality is an extremely tricky and rocky road - you have to get it working in all browsers and then handle firewalls and proxies killing keep-alive connections etc... Why re-invent the wheel. Also, it has a similarly low footprint of less than 10K so it should suit if that is a priority for you.
Both have diferent requirements and address diferent scenarios.
If you need realtime updates, like in an online chat, push is a must.
But, if the refresh period is big, as it is in your case (5 minutes), then pool is the appropriate solution. Push, in this case, will require a lot of resource from both the client and the server.
Tip! try to make the page that checks the pool fast and clean, so it doesn't consumes a lot of resources in the server in each request. What I usually do is to keep a flag in memory (like in a session variable) that says if the pool is empty or not... so, I only do havy look in the pool only if it is not empty. When the pool is empty, which is most of the time, the page request runs extremely fast.
Because using a push requires an open HTTP connection to be maintained between your server and each client, I'd go for poll as well - not only is that going to consume a lot of server resources but it's also going to be significantly more tricky to implement as matt b mentioned.
My experience with polling is that if you have a frequent enough polling interval on a busy enough site your web server logs can get flooded with poll requests real quickly.
Edit (2017): I'd say your choices are now are between websockets and long polling (mentioned in another answer). Sounds like long polling might be the right choice based on the way the question mentions that the notifications don't need to be received in real time, an infrequent polling period would be pretty easy to implement and shouldn't be very taxing on your server. Websockets are cool and a great choice for many applications these days, sounds like that might be overkill in this case though.
I would implement a poll just because it sounds simpler to write, and keeping it simple is very valuable.
Not sure if you have taken a look at some of the COMET implementations out there (is that what you mean by AJAX push).
If the user is surfing the site, won't that in effect be requesting information from the server that this notification can piggy-back on?
It's impossible to say whether polling will be more expensive then pushing without knowing how many clients you'll have. I'd recommend polling because:
It sounds like you want to update data about once per minute. Unless notifications are able to arrive at a much faster rate than that, pushing would mean you're keeping an HTTP connection open but seeing very little activity on it.
Polling is built on top of existing HTTP conventions, so any server that talks to web browsers is already ready to respond to ordinary Ajax requests. A Comet– or Flash socket–based solution has different requirements; you'll need something like cometd on the server side and a client-side library that groks server-side push.
So if you needed something heavy-duty to manage a torrent of data and a crapload of clients, I'd recommend Comet. But that doesn't seem to be the case.
There's now a service http://pusherapp.com that is trying to solve this problem once and for all, in a blink. Might be worth checking out. (disclaimer: i am in no way associated with them).
I haven't tried it myself, but some say COMET works and is easier than you think. There's also a Ruby on Rails plug-in called Juggernaut that I've heard talked about highly. Again, I haven't used it, so YMMV, but my understanding is that it takes far fewer resources compared to polling. I believe (can someone confirm?) that COMET is how MacRumorsLive.com delivers live blogging of WWDC Stevenotes.