Actually, I don't think that even this patch can guarantee solid throughput. All I/O (including SMSc) is handled by a polling thread. AFAIK bb_smscconn_sent just puts the sms in the connection queue, where the connection layer takes over and sends it. If there is a queue for that SMSc, the connection layer will still try to empty it as fast as it can.
IMHO, any robust throughput should be implemented in the connection layer. BR, Nikos ----- Original Message ----- From: Nikos Balkanas To: Donald Jackson ; Damian Viano Cc: [email protected] Sent: Wednesday, June 10, 2009 9:51 AM Subject: Re: [PATCH] Guaranteed throughput smsc-independent Hi Donald, I think you hit the nail exactly on the head. AFAIK, there is a single thread at this point, sms_router, that is implementing bb_smscconn_sent. Therefore all sms in the outgoing smscs should be delayed. Kannel doesn't raise threads on demand. There may be an additional issue. This approach, which delays everything by a set amount if there is a throughput constraint, will utilize exactly half the available bandwidth. If for example an smsc has a throughput of 10 sms/1', and there is a queue of 20 sms, In the old way, assuming it works correctly, the first 10 should be send at once at time 0, and the rest at atime 1' with an average delay of 30"/sms. With this approach, 1 sms would be sent every 6", with an average delay of 1'/sms. Coupled with the previous point, it could spell problems for queues in large installations that use many smscs to load balance. The advantage is that it offers solid throughput handling. I am a wap person, so I leave this decision for peoaple with a lot of smss. BR, Nikos ----- Original Message ----- From: Donald Jackson To: Damian Viano Cc: [email protected] Sent: Wednesday, June 10, 2009 12:06 AM Subject: Re: [PATCH] Guaranteed throughput smsc-independent Ignore my last mail! I see these will be called per thread :) 2009/6/9 Donald Jackson <[email protected]> I will review in more detail when I get a moment, my concern here is that you are now sleeping at the bearerbox layer which could effect delivery to other SMSC's. The current throughput limits although the logic is per-smsc, they sleep within their own threads as to not delay any others. Have you addressed this scenario? 2009/6/9 Damiαn Viano <[email protected]> Hi list: I've seen kannel not respect throughput at all, at least with the fakesmsc and looking around find the following bug: http://redmine.kannel.org/issues/show/332 With the following patch attached: http://redmine.kannel.org/attachments/104/0000332-emi_patch_ack_v3.txt Inspired by that one, I've implemented a smsc-independent throughput patch. The idea is to enforce the throughput from the beaberbox side instead of having to implement the same login in every smsc. This is possible due to the bb_smscconn_sent() and bb_smscconn_send_failed() callbacks from the smscs implementation. They MUST call one of this callbacks after sending a message either successfully(_sent) or with failure(_failed), so we can make them sleep there, making sure they NEVER go over the configured throughput. There's only one downside to this smsc-independent approach which is that we don't, and can't (without cluttering the interface, AFAIK) know how much time the smsc takes in actually sending the sms, so we assume it takes nothing, this would, practically give us a somewhat smaller real throughput, but I though that's better than the previous behaviour (which for me, flooded the smsc). I've only tested this with fakesmsc so far, and only commented out the previous throughput implementation in that smsc, doing the rest is trivial and I can do it (or anyone else can), but this first iteration is to gather opinions about this approach. Once again the patch is against the current stable release, I can update it if needed, just let me know. Again I would love comments, questions, commits, rants, whatever :) For reference: diffstat kannel-1.4.3-throughput.patch bb_smscconn.c | 46 ++++++++++++++++++++++++++++++++++++++++++++++ smsc/smsc_fake.c | 9 +++++++-- smscconn_p.h | 2 ++ 3 files changed, 55 insertions(+), 2 deletions(-) Hope to help. Damiαn Viano(Des). P.D.: Also there's an info line in smsc_fake.c to count the number of sms, which should be removed from the final version, is only there for debugging purposes. -- Donald Jackson http://www.ddj.co.za/ donaldjster(a)gmail.com -- Donald Jackson http://www.ddj.co.za/ donaldjster(a)gmail.com
