Re: [Serusers] still no help - usrloc synchronization

MessageAFAIK, Altheon, Cisco, and F5 all use NAT. There seems to be a general understanding that tunneling or direct routing has better scalability, but I have found many statements where the conclusion is that: LVS DR, Tun, and NAT will have the same maximum throughput, but that LVS NAT will utilize more CPU and memory for the same load and you thus need a better configured server. Ref: http://www.ntua.gr/lvsp/Joseph.Mack/HOWTO/LVS-HOWTO.LVS-NAT.html#lvs_nat_pe… g-) ----- Original Message ----- From: Matt Schulte To: kramarv(a)yahoo.com ; Java Rockx Cc: serusers(a)lists.iptel.org Sent: Tuesday, April 12, 2005 02:31 PM Subject: RE: [Serusers] still no help - usrloc synchronization Forgive my ignorance, why is NAT a bottleneck? Are you refering to CPU load or managibility? With alot of load balancers, NAT is the only way to go. -----Original Message----- From: Tina [mailto:kramarv@yahoo.com] Sent: Monday, April 11, 2005 11:37 AM To: Java Rockx Cc: serusers(a)lists.iptel.org Subject: Re: [Serusers] still no help - usrloc synchronization Pretty good news, Paul. I've only tested IP tunneling. It worked for me. The problem I described was born in my thoughts. I've also read some complaints from SIP-LVS users. The only configuration where Call-Id stickness does not necessary is LVS-NAT. Unfortunately, NAT becomes bottleneck very fast. Java Rockx <javarockx(a)gmail.com> wrote: Tina, I wish I had more information on that, but that LVS stuff is a black art to me. One of our engineers whipped up an LVS configuration that seems quite happy with the Call-ID "stickness" you described. All I can say right now is that it is apparently possible to do this with 100% open source. (i'm keeping my fingers crossed) :-) Regards, Paul On Apr 10, 2005 11:03 PM, Tina <kramarv(a)yahoo.com> wrote: Paul, thanks a lot, it helps a lot. The thing I do not understand - how you are going to solve call-id stickness (see my another post). Since we're also use LVS I had to find something proprietary, I am afraid this is not a best solution, but the only I have today. KRs, Tina Java Rockx <javarockx(a)gmail.com> wrote: Tina, I really don't know how the LVS server is configured because our network guy that we acquired from RedHat set all that stuff up. I do know that we basically have this set up: I hope the formatting goes well :-) +-----------+ +----------+ +-------+ +---------+ +--------------------+ | internet |------| Cisco |------| LVS |-----| Cisco |----| Application | | cloud | | 3600 | | | | 3600 | | & DB Servers | +-----------+ +----------+ +--------+ +---------+ +--------------------+ So the "Application & DB Servers" box represents multiple servers (all are dedicated to their service). These include SER, MySQL, Apache, configuration management, monitoring, RTP proxies, etc. MySQL is active-active so we have two-way replication. We only have 2 MySQL servers. And this is all we plan on ever having as MySQL is more than capable of handling anything ser can throw at it (given the right hardware). Each ser server is 100% oblivious to other SER servers except for when handling REGISTER messages. Here is basically the ser.cfg for each SER server: listen=10.3.0.221 # this will be 10.3.0.222 on the peer sip proxy modparam("usrloc", "db_mode", 1) # write-through route { if (method=="REGISTER") { if (src_ip==10.3.0.221) { # ip of peer ser proxy save_memory("location"); break; } else { save("location"); t_replicate("10.3.0.222"); break; }; }; } We have identifed a deficiency in the usrloc module urecord.c file whereby when using write-though mode the save_memory() function still attempt to update MySQL. This is a very bad thing because it causes a primary key violation on the location table. Yesterday I posted a patch to serdev to fix this, but today we identified another case where save_memory() tried to insert in to the location table. I'll post a revised patch to serdev as soon as I get a chance. So here is what happens. Assume I have two sip routers and two MySQL servers as: sip-01 is 10.3.0.221 sip-02 is 10.3.0.222 db-01 is 10.2.0.21 db-02 is 10.2.0.22 Now let's just assume that sip-01 is connected to db-01 and sip-02 is connected to db-02. When a SIP UA connects to a sip router (via LVS), assume sip-01 in this example and REGISTERs, sip-01 will call save("location") which updates/inserts to the location table. sip-01 then calls t_replicate to sip-02. sip-02 recieves the REGISTER message and calls save_memory("location") to update its in-memory cache only. It should never hit db-02. Now db-01 eventually (within a second or two) replicates to db-02 and now sip-02 is good to go, even in the event of a restart. Our difficulties mostly surrounded the fact th at ser was inserting the physical IP of eth0 on the server as the top VIA rather than the VIP as required. We solved this by using record_route_preset() and passing the VIP as the IP. We also had to bring up a dummy interface in order to get ser to honor the VIP. Hope this helps. Regards, Paul On Apr 8, 2005 7:19 PM, Tina <kramarv(a)yahoo.com> wrote: Paul, we are using LVS as well. It creates template connection according to source Ip (e.g. UAC) and forwards to one of the tunneled IP. I tested this a couple of weeks ago, ser was happy and inserted VIP into via header, the only issue is to route the calls into right server. I agree, there is no such thing as SIP-level clustering. However, how would you solve location problem by DB replication - 1) usrloc does not update from DB....2) even though, I am afraid it can become bottleneck and lock down users ... KRs, Tina Java Rockx <javarockx(a)gmail.com> wrote: See my inline comments. Regards, Paul On Apr 8, 2005 1:08 PM, Greger V. Teigre <greger(a)teigre.com> wrote: Hi Tina, Thanks :-) I believe the load balancer can be in front and transparantly translate addresses in the SIP messages: User agent always sees a.b.c.d | Load balancer with public IP a.b.c.d | | ser1 (10.0.0.2) ser2 (10.0.0.3) Paul: You have a similar load-balancing solution, right? Do you use a commerical load balancer that you can reveal the name of? :-) Our platform is 100% open source. Our load balancer is LVS. I have no idea how it actually works. One if our engineers from RedHat set all that networking stuff up and it's way beyond me. The load balancer need to understand SIP (sort of) and use something (ex. IP address of user agent) to select SER server to send to and rewrite incoming to 10.0.0.2 and 10.0.0.3 respectively and outgoing to a.b.c.d. ser1 and ser2 will happily believe they are alone and communicate directly with user agent using their own private addresses. Both should know about all users (usrloc). Paul just recently posted a small patch for a setup where they use two-way replication between two That small patch will need a slight revision. We discovered through further testing that that patch only fixed attempts to INSERT into the location table. We didn't cover the UPDATE route, unfortunately. I believe Andrew has that patch done and is testing it. So I'll post it to serdev as soon as we know it works - possibly later today. mysql servers and where each SER instance will save all REGISTERs it receives from user agents, but only save to memory those replicated from the other SER peer. Klaus pointed out that the same can be done by not doing mysql replication, but have one DB for each SER and just use t_replicate() and save_noreply() (for two servers, for more than two you need to use forward_tcp()). I have heard complaints that this approach will generate too much network traffic. Also, t_replicate() does not guarantee delivery of the replication (forward_tcp is better, but not fool proof). So, the "best" solution would probably be to hope that Paul's I fully agree. Replication at the sip proxy level, IMHO, is not a good idea. From a purist point of view, replication is not a function of a SIP router - nor should it be. It is a function of the database and we believe that the MySQL replication is a better approach from a speed and reliabilty point of view - not to mention that it simplifies the tasks of the sip router. recent TCP/UDP patch for network access to FIFO commands get included in the CVS. Then, a new replication module should use XMLRPC or whatever protocol the patch is using to forward (in a guaranteed mode) the usrloc information, which then would be injected directly into SER's memory, just like save() would, and of course in whatever db_mode you would like. The replicaton module would have to keep a queue (of course, flushed to a DB or something) in case of network problems, so that the replication would be done (sooner or later) regardless of problems. Of course, this is a major undertaking, so for most replication scenarioes, I would think the forward_tcp() + save_noreply() would work just fine. g-) _______________________________________________ Serusers mailing list serusers(a)lists.iptel.org http://lists.iptel.org/mailman/listinfo/serusers -------------------------------------------------------------------- Do you Yahoo!? Make Yahoo! your home page __________________________________________________ Do You Yahoo!? Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com __________________________________________________ Do You Yahoo!? Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com ------------------------------------------------------------------------------ _______________________________________________ Serusers mailing list serusers(a)lists.iptel.org http://lists.iptel.org/mailman/listinfo/serusers