enyim.com Memcached Client

New Post: Struggling to configure basic Memcache

tim3452 — Mon, 18 Aug 2008 16:53:58 GMT

Or, another way to ask about scaling on a single machine:
I see that the sample.config has this <servers> node:

But what do I run from the command line to create this setup? I had assumed something like so:

memcached -d install
memcached -d start -p 20002
memcached -d start -p 20004
memcached -d start -p 20006
memcached -d start -p 20008

But that doesn't work - Data simply won't be persisted or retrieved.

If I install it only on one other port (like 20002), then it works. But I cannot install it on multiple ports (on the same machine). The SetCache runs, I can see that activity is occuring (from running a port listener), but nothing is ever returned in the GetCache.

Also, only one MemCache service instance is running. Not sure if this is supposed to work by having multiple MemCache services running - one for each port.

New Post: Struggling to configure basic Memcache

tim3452 — Fri, 15 Aug 2008 22:10:30 GMT

Hello,
I was excitec when I learned about Enyim and Memcache, but I cannot find any tutorial or walkthrough to explain it. In particular, I'm really struggling to:

1. Increase the memory size of the memcached process (it defaults to 64MB, but and I try using the -m switch to change it, but that either has no affect (it caps at 68076). I've tried things like stopping the service ("memcached -d stop") and then restarting: "memcached -d start -m 256"

2. Can I install multiple processes on a single server? The Memcached home page hinted at this right at the top (http://www.danga.com/memcached/), "memcached -d -m 2048 -l 10.0.0.40 -p 11211". It seems something like setting it up on additional ports? But again, I try variations of
"memcached -d start -p 11222" to create a second service, but I see nothing happening. Even if there was a way to increase the process size beyond 64MB, it seems you'd want multiple instances of the process in order to provide the biggest cache.

3. I hear references to this "deadtime" (once you start the server, how long it takes to become available). How can I set that? I don't see a config switch for it.

4. If I set the "-v" switch, where does it log output too? I've got Log4Net installed, but I can't find any created log files.

I'd appreciate anyone's insight. I figure these are the most basic scenarios, so they would need to have been solved already.

Thanks

New Post: Connection count?

damoh — Mon, 28 Jul 2008 23:01:10 GMT

I am wondering what the best usage pattern is with the MemcachedClient is too.

Should we have a single instance (static or otherwise) of a MemcachedClient or should we create an instance each time we use it, i.e.

using(MemcachedClient cache = new MemcachedClient()){
....
}

I suspect we should create an instance each time as there may be issues accessing a MemcachedClient object with multiple threads?

Created Issue: Performance counters

andybrummer — Thu, 17 Jul 2008 02:56:57 GMT

It would be nice if there were performance counters for things like reads and writes per second. Number of connections in the pool, number of active nodes, killed and resurected nodes and the like. I don't have a full list right now, but I'd be able to work something up if you were interested, along with contributing to the instrumentation.

Commented Issue: asynchronous methods

andybrummer — Thu, 17 Jul 2008 02:52:19 GMT

Since this is a network I/O based system it would be nice if asynchronous methods were available (Begin*/End* and *Async) that could use IOCP under the hood (Socket.BeginSend/BeginRead). I may have time to contribute this if you're interested...
Comments: ** Comment from web user: andybrummer **

I'm very interested in this as well, and would be interested in contributing as well.

Commented Issue: Unbalanced distribution of cache items per server

andybrummer — Thu, 17 Jul 2008 02:50:46 GMT

Hi, we are using successfully in our production environment your memcached client.
After 2 weeks of continuous working we have experienced an unbalanced distribution of items per server.
It seems that some of the servers are more loaded than others, visibly the first and the last according to the order in which they appear in the configuration.

These are our statistics:

SERVERS CURRENT_ITEMS
srv01:11211 120681
srv01:11212 61299
srv01:11213 35360
srv01:11214 38134
srv02:11211 70774
srv02:11212 57018
srv02:11213 74132
srv02:11214 71798
srv03:11211 34891
srv03:11212 76513
srv03:11213 66546
srv03:11214 118872
Comments: ** Comment from web user: andybrummer **

Sorry, just re-read the comment. I'd like to add that the value after the second colon is the percentage of the hash space covered by the node, so the first node is getting 60% of the hashes.

Commented Issue: Unbalanced distribution of cache items per server

andybrummer — Thu, 17 Jul 2008 02:49:07 GMT

I have some additional data for this issue:

4 servers:
192.168.1.95:1121
192.168.1.101:1121
192.168.1.98:1121
192.168.1.99:1121

I added some code to the defaultNodeLocator file to generate the hash space coverage for the nodes and this is what I got

192.168.1.98:11211:60.3270117799582
192.168.1.99:11211:7.26672550832544
192.168.1.101:11211:12.2046676725626
192.168.1.95:11211:10.1610983279909

So 60 of all the hashes map to the first server. I increased the ServerAddressMutations to somewhere around 1000 and the distribution was better. I think the algorithm might need some tweaking to make sure the distribution is close to even.

I just set a breakpoint on line 69 and look at the contents of the builder.

Commented Issue: .Stats() does not resurect servers

ikatkov — Tue, 15 Jul 2008 13:52:26 GMT

.Stats() does not resurect servers
Comments: ** Comment from web user: ikatkov **

So, if one server was marked as Dead it will be dead until somebody call GET/SET... command. If all you do is calling Stats() this server will be dead forever. Workaround - always call dummy GET before .Stats() or recreate Memcached client

Created Issue: .Stats() does not resurect servers

ikatkov — Tue, 15 Jul 2008 13:50:19 GMT

.Stats() does not resurect servers

Created Issue: GET("") fails with System.NotSupportedException

ikatkov — Tue, 15 Jul 2008 13:47:32 GMT

GET("") fails with System.NotSupportedException

Created Issue: change behavior of default keytransformer to encode forbidden chars

jconley — Fri, 11 Jul 2008 18:20:37 GMT

It would be quite helpful if the default key transformer simply encoded bad characters into %HEX or some other encoding mechanism, rather than throwing an error.

New Post: which keytransformer to use?

jconley — Fri, 11 Jul 2008 17:34:06 GMT

We're creating a framework that exposes a caching layer and don't want to limit the characters that can be in the key. How reliable are TigerHash/SHA1 in terms of hash collisions in the real world? I noticed the default for the library is just to pass through the keys. We can use Base64, but it seems like a waste of bytes. :)

New Post: Node Location incompatible with PHP PECL memcache lib 'consistient'?

davidljung — Wed, 09 Jul 2008 04:15:31 GMT

Hi all.
I'm in the process of implementing a cache for a system with a mix of PHP and C# .NET code. For the PHP end (web-server code), I'm using the standard PHP PECL memcache library. It isn't as configurable as this client lib, but it has an option named "memcache.hash_strategy" which can take the values "standard" or "consistient". I've set it to "consistient". There is also an option named "memcache.hash_function" which can take the values "crc32" or "fnv" and which I've set to "fnv". The docs say ( http://us2.php.net/manual/en/memcache.ini.php ):
"Controls which hash function to apply when mapping keys to servers, crc32 uses the standard CRC32 hash while fnv uses FNV-1a."

If I set the PHP strategy to "standard" everything works as I'd expect (with limited testing). However, I wish to use the "consistient" option.

On the C# .NET end, I'm using the default NodeLocator and KeyTransformer, which I understand uses the FNV1a has class and doesn't transform the keys (except for disallowing invalid chars). I'm using a custom TransCoder to match what I have on the PHP end (encoded JSON strings).

Now, with one memcached server everything works as expected. However, if I have two memcached servers, the key->node mapping between PHP and this client isn't the same. For example, the key "ac_mykey12" is mapped to the first server by the enyim client and to the second by the PHP client.
In both cases, the same two servers are added using addServer() in the same order (as IP address strings).

Based on the configurations I'd guess the key hashing it the same. How does the "consistient" PHP setting effect the node selection? (I understand what it does functionally - keep the same keys mapping to the same nodes regardless of their server index - but how? is the IP address being incorporated into the hash? if so, exactly how?)

I guess my real question is: how can I duplicate the behaviour of PHP's "consistient" stragety with enyim? (is it implemented already, or do I need to implement my own NodeLocation? If so, I don't know what PHP is doing...) Has anyone achieved this already?

Any ideas/suggestions/pointers appreciated.
Thanks.

New Post: ReaderWriter perf

onovotny — Tue, 08 Jul 2008 17:23:09 GMT

I just wanted to suggest switching over to the ReaderWriterLockSlim lock in the .NET 3.5 library instead of using the ReaderWriterLock. The perf of the new lock is significantly better and they've fixed issues with the upgrade capability where you have to check the lock cookie.

Also, as another perf optimization, you could probably use a new object expression compiled into a delegate for the fast activator instead of creating a new assembly. That'll allow for code to be cleaned up instead of requiring an appdomain unload.

Just a couple of thoughts.

Commented Issue: Un-seal IMemcachedKeyTransformer Implementations

jconley — Thu, 03 Jul 2008 17:36:33 GMT

We'd like to extend the transformers for the purpose of automatically adding a prefix to every key so we can share our memcached farm between applications without worrying about key collisions. To do this today we have had to copy your transformer implementations.
Comments: ** Comment from web user: jconley **

I should say, it would also serve our purpose if there were a keyPrefix configuration item, as that's what we've used this for anyway.

Created Issue: Un-seal IMemcachedKeyTransformer Implementations

jconley — Thu, 03 Jul 2008 17:35:07 GMT

Commented Issue: does not reconnect to server after losing connection

jconley — Thu, 03 Jul 2008 17:32:28 GMT

1) Successfully use the Enyim memcached client
2) Restart memcache server
3) Notice that the client doesn't bring the restarted server into the pool

Or maybe I didn't wait long enough during my test. Is there a reconnect thread? I am using the latest release build.
Comments: ** Comment from web user: jconley **

this seems to be working well in 1.2.0.2

New Post: Connection count?

trollfan — Thu, 26 Jun 2008 05:35:52 GMT

O, everytime I needed to use the client I was doing

MemcachedClient cache = new MemcachedClient();

and then calling the method I needed. Is this just me being an idiot? Should I have this being static so there's only one instance of the MemcachedClient ?

New Post: Connection count?

trollfan — Thu, 26 Jun 2008 04:43:47 GMT

Hi, I noticed that if I run netstat on the machine running the memcache server, when my web app makes a request to memcache there's a heck of a lot of ESTABLISHED connections coming over and hitting the port memcache is running on. Is this normal? Or can you kind of explain why it is doing this? It seems like there's dozens (maybe hundred) connections happening for just one Get command. It's been a while since I had to know the details of sockets and what not.

Should there be a normal ratio between the GetCount and ConnectionCount or any of the stats?

Thanks!

New Post: client throws some exceptions

fby001_wangdui — Wed, 25 Jun 2008 07:49:51 GMT

when i use JMeter to test my website, i got some exceptions:

2008-06-25 15:15:45,968 [138] ERROR Enyim.Caching.Memcached.MemcachedNode+InternalPoolImpl [(null)] - Could not init pool.
System.Net.Sockets.SocketException: 由于系统缓冲区空间不足或队列已满，不能执行套接字上的操作。 127.0.0.1:11211
   在 System.Net.Sockets.Socket.DoConnect(EndPoint endPointSnapshot, SocketAddress socketAddress)
   在 System.Net.Sockets.Socket.Connect(EndPoint remoteEP)
   在 Enyim.Caching.Memcached.PooledSocket..ctor(IPEndPoint endpoint, TimeSpan connectionTimeout, TimeSpan receiveTimeout, Action`1 cleanupCallback)
   在 Enyim.Caching.Memcached.MemcachedNode.InternalPoolImpl.CreateSocket()
   在 Enyim.Caching.Memcached.MemcachedNode.InternalPoolImpl.InitPool()
2008-06-25 15:15:45,968 [138] WARN Enyim.Caching.Memcached.MemcachedNode+InternalPoolImpl [(null)] - Marking pool 127.0.0.1:11211 as dead

2008-06-25 15:17:04,437 [138] ERROR Enyim.Caching.Memcached.MemcachedNode+InternalPoolImpl [(null)] - Could not init pool.
System.Net.Sockets.SocketException: 通常每个套接字地址(协议/网络地址/端口)只允许使用一次。 127.0.0.1:11211
   在 System.Net.Sockets.Socket.DoConnect(EndPoint endPointSnapshot, SocketAddress socketAddress)
   在 System.Net.Sockets.Socket.Connect(EndPoint remoteEP)
   在 Enyim.Caching.Memcached.PooledSocket..ctor(IPEndPoint endpoint, TimeSpan connectionTimeout, TimeSpan receiveTimeout, Action`1 cleanupCallback)
   在 Enyim.Caching.Memcached.MemcachedNode.InternalPoolImpl.CreateSocket()
   在 Enyim.Caching.Memcached.MemcachedNode.InternalPoolImpl.InitPool()

"通常每个套接字地址(协议/网络地址/端口)只允许使用一次" means: usually each socket can be used once.
"由于系统缓冲区空间不足或队列已满，不能执行套接字上的操作。" means: system buffer or queue is full, the operation on socket can not be done.

the configuration is:
   ...
   <servers>
    <add address="127.0.0.1" port="11211" />
   </servers>
   <socketPool minPoolSize="300" maxPoolSize="1000" connectionTimeout="00:00:10" deadTimeout="00:02:00" />
   ...

JMeter's configuration is:
50 threads for accessing my website

this happens when i use the latest version (1.2.0.2) client.i guess socket pool has some errors maybe ,for example: have not close the socket connection?