I need to write a socket server that will be handling at least something about 1000 (much more in future) low-traffic permanent connections. I have made a draft version on PHP for testing purposes (we are developing a monitoring hardware, so we needed to develop and test a conversation protocol and hardware capabilities), which suited me very well when i had just a couple of clients connected. But when the amount of connections had grown to ten, some critical issues appeared. Here some info about server architecture:
I have a master process, which waits for socket connections and on connecting creates a child process (that serves this connection from now on) using pcntl_fork(). Also i am setting up a PDO connection to MySQL in master process. All the child processes are sharing the same single PDO object. At first i was afraid of getting some collisions during simultaneous queries, but i haven't encountered them, even through stress-test (10 children were making queries in the loop without stopping). But there is usleep(500000) in each child, so it could be luck, though i had this testing running for a couple of hours. But such load should not be present even at 1k clients connected, due to rare conversations between them and server.
So here is my first question: is it safe to use single PDO object for a big amount of child processes (ideally there would be around 1000)? I can use single connection for each child, but MySQL doesn't support nearly as much connections.
The second issue is in getting parasite MySQL connections. As i mentioned before, i have only one PDO object. But when i have more than one clients connected, and after they had run some queries, i see in mytop that there is more than one DB connection, and i could not find any correlation between the amount of connections and amount of child processes i have. For example i have 3 childs, and 5 DB connections. I tried to establish persistent connections, and it didn't changed anything.
Second question: Is it PDO who makes those additional connecitons to MySQL, or it is the MySQL driver? And is there a way to force them to use one connection? I don't think it could be my fault, my code prints an alert to console every time i call method which creates PDO object, and that happens only once, at the script start, before forking. After that i only run querys out of children, using parent's PDO object. Once again i can not afford to have so many connections due to MySQL limitations.
Third question: Will be one thousand of socket connections a problem by itself? Aside of the CPU and database load, i mean. Or i should do some amount of lesser servers (128 connections for example), that will tell the clients to connect to other one if max number of connects is exceeded?
Thanks in advance for your time and possible answers.
Currently your primary concern should be your socket server architecture. Forking a process for each client is super heavy. AFAIK an average PC can tolerate around 2000 threads and it's not going to work fast. Switching between processes means that CPU should save its state in memory, and if you have enormous amount of processes, CPU will be busy with memory IO and will have little time for actually doing stuff.
You may want to look at Apache for inspiration. In Apache they use a fixed amount of worker processes/threads, each process/thread working with multiple clients via select function and sockets in a non-blocking mode. This is a far more robust approach.
Regarding database IO, I would spawn a process/thread that would be the sole owner of database connections. Worker processes would communicate with the DB IO process using IPC (in case of processes) or lock-free queues (in case of threads). This approach makes you independent of PDO implementation details (if it is thread safe or does it spawn connections etc).
P. S. I suspect that you actually spawn new PDO objects with forking (forking merely means making a copy of a process with its memory and everything inside it) and PDO objects create and shut down connections on demand. It may explain why you're not seeing correlation between low traffic clients and DB connections.
Related
I wanted to know if I forked a processes after I bonded the IP and port of the server. Will the fork be able to accept the connection. To extend on that if I have 10 forks running all trying to accept is there a chance that more then 1 could accept the same connection or is there some locking on that?
A few days ago I felt like writing writing a http server in php. So it can handle more then one connection at a time. The master processes accepted the connection reads the data in and passes it to a thread via a unix socket. So far on my laptop i can get 1000 connections a second on a little page that gives the current date and time. One of the bottle necks is the master processes. Originally i would have loved to gotten the file descriptors of the connections and passed those to the sockets and have them read the data in then processes it.
Yes, the forked children will be able to accept new connections on the same (inherited) listening socket.
Assuming you're using a blocking socket_accept() in all your child processes, you shouldn't experience any performance issues even if it ramps up to 100 processes; the operating system will wake up one child process to handle the connection.
It should be mentioned that it's good practice to benchmark it, using ab or similar load generator tools.
I have a mongodb server in production serving on an EC2 instance. According to the mongodb official documentation, persistent DB connections should ALWAYS be used in production. I've been experimenting with about 50 persistent connections and was getting frequent connection errors (approx 33% of the time) while testing. I'm using this code:
$pid = 'db_'.rand(1,50);
$mongo = new Mongo("mongodb://{$user}:{$pass}#{$host}", array('persist' => $pid) );
Some background on the application, it's a link tracking application that is still ramping up - and is in the range of 500 - 1k writes per hour, nothing too crazy... yet.
I'm wondering if I simply need to allow more persistent connections? How does one determine the right balance of persistent connections versus server resources available?
Thanks in advance everyone.
The persist value is no longer supported as of the most recent driver (1.2.0).
Truth is, it was never really clear what it did in typical Apache+PHP setups. There are several comments on the Google Groups and elsewhere asking for detail, but I did not any evidence that persist or persistent was ever tested with any depth.
Instead, it's all been replaced by connection pooling "out of the box". The connection pooling has obviously been through some changes within the 1.2 line with the addition of the MongoPool class.
There is still no detailed explanation of how the pooling works with Apache, but at least you don't have to worry about persist.
Now despite all of this mess, I have handled 1000 times that traffic on a single MongoDB server via the PHP driver without lots of connection problems.
Are you catching the exceptions?
Can you provide more details about the exact exception?
There may be a code solution.
Are you opening a new connection for each PHP page request, or using a connection pool with 50 persistent connections? If you're opening a new connection each time then you might be quickly running out of resources.
Each connection uses an additional thread on the server, so you could be hitting a limit on the number of threads of network connections, check your server logs in /var/lib/mongodb for errors.
If you're using the official MongoDB PHP driver, then as far as I know it should handle connection pooling for you automatically. If you're connecting to Mongo from 50 separate clients, then consider putting a queue in front of Mongo to buffer the writes.
http://php.net/manual/en/mongo.connecting.php
without Persistent Connections x1000
It takes approximately 18 seconds to execute
Persistent
...it takes less than .02 seconds
In the Java world it is pretty standard for app servers to pool "expensive" resources, like DB connections. On the other hand in dynamic languages, most stacks have little to do with pooled resources and especially DB connections.
E.g. for the popular PHP+MySQL combo, I've rarely seen it used with persistent connection, which can be considered poor-mans pooled connections.
If the concept of pooling DB connections is not that widely implemented, does this mean that the performance/scalability gains might not be all that important, in real-life deployments?
The main reason for connection pooling is the overhead in establishing the connection in the first instance. I have seen this take up to 0.5 seconds in the past.
In a high transactional environment, the ability to keep a connection open, and send multiple requests down the connection, one after the other have significant savings. Therefore, you may not see the gains in a low transactional database, but your application is not going to scale as well, if you ignore this useful pattern.
It also helps to managed the number of open connections in a much clearer way.
If the concept of pooling DB
connections is not that widely
implemented, does this mean that the
performance/scalability gains might
not be all that important, in
real-life deployments?
It just means that there's no built-in connection pool for PHP. Doesn't mean you can't use one, though, for instance with Postgres you can use PGPool.
My general opinion is that connection pools generally aren't used in environments where speed isn't considered as important as other concerns. AOLServer, for example, uses a dynamic language (Tcl) but is highly performant and uses connection pools.
Connection polling is often used because some database vendors limit the number of connections that you have to a given database depending on your license. Open Source databases do not have such limits because they are free. So it is not much of a problem for MySQL.
Another reason to use connection polling is to limit the number of current connections to the database server, as each new connection consumes a lot of memory and you do not want to exhaust your servers memory.
The problem with persistent connections is that they never close until the client processes die. Client processes are in reality the Web server processes handling PHP requests. So, if you configure your Web server to limit the number of simultaneous requests, you also limit the number of opened persistent database connections. You can do that in Apache setting the MaxClients parameter to a reasonable value that does not exhaust your server RAM.
BTW, it would also be wise to move all your static content (CSS, JavaScript, images, etc..) to a separate multithreaded Web server (Nginx, lighttpd, etc..) so the simultaneous user accesses do not make your Apache fork to much processes.
In PDO, a connection can be made persistent using the PDO::ATTR_PERSISTENT attribute. According to the php manual -
Persistent connections are not closed at the end of the script, but
are cached and re-used when another script requests a connection using
the same credentials. The persistent connection cache allows you to
avoid the overhead of establishing a new connection every time a
script needs to talk to a database, resulting in a faster web
application.
The manual also recommends not to use persistent connection while using PDO ODBC driver, because it may hamper the ODBC Connection Pooling process.
So apparently there seems to be no drawbacks of using persistent connection in PDO, except in the last case. However., I would like to know if there is any other disadvantages of using this mechanism, i.e., a situation where this mechanism results in performance degradation or something like that.
Please be sure to read this answer below, which details ways to mitigate the problems outlined here.
The same drawbacks exist using PDO as with any other PHP database interface that does persistent connections: if your script terminates unexpectedly in the middle of database operations, the next request that gets the left over connection will pick up where the dead script left off. The connection is held open at the process manager level (Apache for mod_php, the current FastCGI process if you're using FastCGI, etc), not at the PHP level, and PHP doesn't tell the parent process to let the connection die when the script terminates abnormally.
If the dead script locked tables, those tables will remain locked until the connection dies or the next script that gets the connection unlocks the tables itself.
If the dead script was in the middle of a transaction, that can block a multitude of tables until the deadlock timer kicks in, and even then, the deadlock timer can kill the newer request instead of the older request that's causing the problem.
If the dead script was in the middle of a transaction, the next script that gets that connection also gets the transaction state. It's very possible (depending on your application design) that the next script might not actually ever try to commit the existing transaction, or will commit when it should not have, or roll back when it should not have.
This is only the tip of the iceberg. It can all be mitigated to an extent by always trying to clean up after a dirty connection on every single script request, but that can be a pain depending on the database. Unless you have identified creating database connections as the one thing that is a bottleneck in your script (this means you've done code profiling using xdebug and/or xhprof), you should not consider persistent connections as a solution to anything.
Further, most modern databases (including PostgreSQL) have their own preferred ways of performing connection pooling that don't have the immediate drawbacks that plain vanilla PHP-based persistent connections do.
To clarify a point, we use persistent connections at my workplace, but not by choice. We were encountering weird connection behavior, where the initial connection from our app server to our database server was taking exactly three seconds, when it should have taken a fraction of a fraction of a second. We think it's a kernel bug. We gave up trying to troubleshoot it because it happened randomly and could not be reproduced on demand, and our outsourced IT didn't have the concrete ability to track it down.
Regardless, when the folks in the warehouse are processing a few hundred incoming parts, and each part is taking three and a half seconds instead of a half second, we had to take action before they kidnapped us all and made us help them. So, we flipped a few bits on in our home-grown ERP/CRM/CMS monstrosity and experienced all of the horrors of persistent connections first-hand. It took us weeks to track down all the subtle little problems and bizarre behavior that happened seemingly at random. It turned out that those once-a-week fatal errors that our users diligently squeezed out of our app were leaving locked tables, abandoned transactions and other unfortunate wonky states.
This sob-story has a point: It broke things that we never expected to break, all in the name of performance. The tradeoff wasn't worth it, and we're eagerly awaiting the day we can switch back to normal connections without a riot from our users.
In response to Charles' problem above,
From : http://www.php.net/manual/en/mysqli.quickstart.connections.php -
A common complain about persistent connections is that their state is
not reset before reuse. For example, open and unfinished transactions
are not automatically rolled back. But also, authorization changes
which happened in the time between putting the connection into the
pool and reusing it are not reflected. This may be seen as an unwanted
side-effect. On the contrary, the name persistent may be understood as
a promise that the state is persisted.
The mysqli extension supports both interpretations of a persistent
connection: state persisted, and state reset before reuse. The default
is reset. Before a persistent connection is reused, the mysqli
extension implicitly calls mysqli_change_user() to reset the state.
The persistent connection appears to the user as if it was just
opened. No artifacts from previous usages are visible.
The mysqli_change_user() function is an expensive operation. For
best performance, users may want to recompile the extension with the
compile flag MYSQLI_NO_CHANGE_USER_ON_PCONNECT being set.
It is left to the user to choose between safe behavior and best
performance. Both are valid optimization goals. For ease of use, the
safe behavior has been made the default at the expense of maximum
performance.
Persistent connections are a good idea only when it takes a (relatively) long time to connect to your database. Nowadays that's almost never the case. The biggest drawback to persistent connections is that it limits the number of users you can have browsing your site: if MySQL is configured to only allow 10 concurrent connections at once then when an 11th person tries to browse your site it won't work for them.
PDO does not manage the persistence. The MySQL driver does. It reuses connections when a) they are available and the host/user/password/database match. If any change then it will not reuse a connection. The best case net effect is that these connections you have will be started and stopped so often because you have different users on the site and making them persistent doesn't do any good.
The key thing to understand about persistent connections is that you should NOT use them in most web applications. They sound enticing but they are dangerous and pretty much useless.
I'm sure there are other threads on this but a persistent connection is dangerous because it persists between requests. If, for example, you lock a table during a request and then fail to unlock then that table is going to stay locked indefinitely. Persistent connections are also pretty much useless for 99% of your apps because you have no way of knowing if the same connection will be used between different requests. Each web thread will have it's own set of persistent connections and you have no way of controlling which thread will handle which requests.
The procedural mysql library of PHP, has a feature whereby subsequent calls to mysql_connect will return the same link, rather than open a different connection (As one might expect). This has nothing to do with persistent connections and is specific to the mysql library. PDO does not exhibit such behaviour
Resource Link : link
In General you could use this as a rough "ruleset"::
YES, use persistent connections, if:
There are only few applications/users accessing the database, i.e.
you will not result in 200 open (but probably idle) connections,
because there are 200 different users shared on the same host.
The database is running on another server that you are accessing over
the network
An (one) application accesses the database very often
NO, don't use persistent connections, if:
Your application only needs to access the database 100 times an hour.
You have many, many webservers accessing one database server
Using persistent connections is considerable faster, especially if you are accessing the database over a network. It doesn't make so much difference if the database is running on the same machine, but it is still a little bit faster. However - as the name says - the connection is persistent, i.e. it stays open, even if it is not used.
The problem with that is, that in "default configuration", MySQL only allows 1000 parallel "open channels". After that, new connections are refused (You can tweak this setting). So if you have - say - 20 Webservers with each 100 Clients on them, and every one of them has just one page access per hour, simple math will show you that you'll need 2000 parallel connections to the database. That won't work.
Ergo: Only use it for applications with lots of requests.
On my tests I had a connection time of over a second to my localhost, thus assuming I should use a persistent connection. Further tests showed it was a problem with 'localhost':
Test results in seconds (measured by php microtime):
hosted web: connectDB: 0.0038912296295166
localhost: connectDB: 1.0214691162109 (over one second: do not use localhost!)
127.0.0.1: connectDB: 0.00097203254699707
Interestingly: The following code is just as fast as using 127.0.0.1:
$host = gethostbyname('localhost');
// echo "<p>$host</p>";
$db = new PDO("mysql:host=$host;dbname=" . DATABASE . ';charset=utf8', $username, $password,
array(PDO::ATTR_EMULATE_PREPARES => false,
PDO::ATTR_ERRMODE => PDO::ERRMODE_EXCEPTION));
Persistent connections should give a sizable performance boost. I disagree with the assement that you should "Avoid" persistence..
It sounds like the complaints above are driven by someone using MyIASM tables and hacking in their own versions of transactions by grabbing table locks.. Well of course you're going to deadlock! Use PDO's beginTransaction() and move your tables over to InnoDB..
seems to me having a persistent connection would eat up more system resources. Maybe a trivial amount, but still...
The explanation for using persistent connections is obviously reducing quantity of connects that are rather costly, despite the fact that they're considerably faster with MySQL compared to other databases.
The very first trouble with persistent connections...
If you are creating 1000's of connections per second you normally don't ensure that it stays open for very long time, but Operation System does. Based on TCP/IP protocol Ports can’t be recycled instantly and also have to invest a while in “FIN” stage waiting before they may be recycled.
The 2nd problem... using a lot of MySQL server connections.
Many people simply don't realize you are able to increase *max_connections* variable and obtain over 100 concurrent connections with MySQL others were beaten by older Linux problems of the inability to convey more than 1024 connections with MySQL.
Allows talk now about why Persistent connections were disabled in mysqli extension. Despite the fact that you can misuse persistent connections and obtain poor performance which was not the main reason. The actual reason is – you can get a lot more issues with it.
Persistent connections were put into PHP throughout occasions of MySQL 3.22/3.23 when MySQL was not so difficult which means you could recycle connections easily with no problems. In later versions quantity of problems however came about – Should you recycle connection that has uncommitted transactions you take into trouble. If you recycle connections with custom character set configurations you’re in danger again, as well as about possibly transformed per session variables.
One trouble with using persistent connections is it does not really scale that well. For those who have 5000 people connected, you'll need 5000 persistent connections. For away the requirement for persistence, you may have the ability to serve 10000 people with similar quantity of connections because they are in a position to share individuals connections when they are not with them.
I was just wondering whether a partial solution would be to have a pool of use-once connections. You could spend time creating a connection pool when the system is at low usage, up to a limit, hand them out and kill them when either they've completed or timed out. In the background you're creating new connections as they're being taken. At worst case this should only be as slow as creating the connection without the pool, assuming that establishing the link is the limiting factor?
Trying to separate out my LAMP application into two servers, one for php and one for mysql. So far the application connects locally through a file socket and works fine.
I'm worried about the number connections I can establish if it is over the network. I have been testing tcp connections on unix for benchmark purposes and I know that you cannot exceed a certain amount of connections per second otherwise it halts due to the lack of resources (be it sockets, or file handles or whatever). I also understand that php does not implement connection pooling so for each page load a new connection over the network must be made. I also looked into pconnect for php and it seems to bring more problems.
I know this is a very very common setup (php+mysql), can anyone provide some typical usage and statistics they get out of their servers? Thanks!
The problem is not related to running out of connections allowed my MySQL. The main problem is that unix cannot very quickly create and tear down tcp connections. Sockets end up in TIME_WAIT and you have to wait for a period before you free up more sockets to connect again. These two screenshots clearly shows this pattern. MySQL does work up to a certain point and then pauses because the web server ran out of sockets. After certain amount of time passed, the web server was able to make new connections.
alt text http://img35.imageshack.us/img35/3809/picture4k.png
alt text http://img35.imageshack.us/img35/4580/picture2uyw.png
I think the limit is at 65535. So you'd have to have 65535 connections at the same time to hit that limit since a regular mysql connection closes automatically.
mysql_connect()
Note: The link to the server will be closed as soon as the execution of the script ends, unless it's closed earlier by explicitly calling mysql_close().
But if you're using a persistent mysql connection, then you can run into trouble.
Using persistent connections can require a bit of tuning of your Apache and MySQL configurations to ensure that you do not exceed the number of connections allowed by MySQL.
Each MySQL connection actually uses several meg of ram for various buffers, and takes a while to set up, which is why MySQL is limited to 100 concurrent open connections by default. You can up that limit, but it's better to spend your time trying to limit concurrent connections, via various methods.
Beware of raising the connection limit too high, as you can run out of memory (which, I believe, crashes mysql), or you may push important things out of memory. e.g. MySQL's performance is highly dependent on the OS automatically caching the data it reads from disk in memory; if you set your connection limit too high, you'll be contending for memory with the cache.
If you don't up your connection limit, you'll run out of connections long before your run out of sockets/file handles/etc. If you do increase your connection limit, you'll run out of RAM long before you run out of sockets/file handles/etc.
Regarding limiting concurrent connections:
Use a connection pooling solution. You're right, there isn't one built in to PHP, but there are plenty of standalone ones out there to choose from. This saves expensive connection setup/tear down time.
Only open database connections when you absolutely need them. In my current project, we automatically open a database connection when the first query is issued, and not a moment before; we also release the connection after we've done all our database work, but before the page's HTML is actually generated. The shorter the period of time you hold connections open, the fewer connections will be open simultaneously.
Cache what you can in a lighter-weight solution like memcached. My current project temporarily caches pages displayed to anonymous users (since every anonymous user gets the same HTML, in the end -- why bother running the same database queries all over again a few scant milliseconds later?), meaning no database connection is necessary at all. This is especially useful for bursts of anonymous traffic, like a front-page digg.