Yes, Vincent Bernat blog posts is really good. However, using these softwares on EC2 which are VMs does not give the same performance in all meanings. But I think that it stills perform pretty good.
I am using taskset for all processes. Haproxy goes to cpu 01 and each stud process gets bound to 02.03.04 etc. depending on how many processes of stud I run with. The latest test I'm doing is with SSL cache turned off on the client side which means that for each new connection it negotiates again. And for each connection it is doing these requests I mentioned before. The result are totally different... Running with negotiation for each connection gives better result using more stud processes. Now I get the worst result using stud with n=3 and the best result with stud n=6 :) Also I get much less actual connection/s in haproxy, only 800 conns/s compared to 2000 conns/s. This was definitely much heavier for the webserver and stud proved to handle it better with more processes. Also HAproxy worked less, it had an idle time of ~50% compared to 15% with the tests before. Swap was not used, it was plenty of memory left. "Last but not least, when benching a platform, it's a bad idea to introduce random stuff. IE your "think time"." - I agree, using a script on the backend would have been much better. The only reason I made the script as is with "wait time" was because it's a customer environment that I am currently working with and putting in a custom script was a longer process then just solving the problem my self :) The script should make use of http keep alive with a small wait time between each request in order to not close the connection. Thanks /E -----Original Message----- From: Baptiste [mailto:bed...@gmail.com] Sent: den 2 november 2011 14:22 To: Erik Torlen Cc: haproxy@formilux.org Subject: Re: Haproxy timing issues Hi Erik, I doubt this could improve things because of virtualization, but have you tried binding processes to CPUs? On a physical hardware, the purpose is to benefit of the l2/l3 CPU cache, mainly for network IO and HAProxy, and also reducing the overhead of the CPU moving processes from a core to an other one. Again, I guess the virtualisation layer would lower the impact of L2/L3 CPU cache :) In your c1.xlarge, you have 8 virtual CPU cores. I guess you wanted 6 studs + 1 HAProxy to get the best of your "hardware". Since your best results are obtained with 3 CPU cores, it seems, you can use only 4 physical CPUs. Read Vincent article: http://vincent.bernat.im/en/blog/2011-ssl-benchmark.html You'll see that stud capacity grows almost linearly with the number of stud process. When running 6 studs, have you recorded the vmstat output? Compare it with a record without any load and maybe compare it too when running 3 studs processes, maybe something obvious will appear. I guess there are some locks somewhere, maybe in the underlying hypervisor... You're maybe hitting the famous virtualization clock drift: http://software.intel.com/en-us/blogs/2009/06/25/virtualization-and-performance-vm-time-drift/ Since HAProxy is event-driven, it is very sensible to clock synchronization. Once your VM is overloaded, all the processes are impacted, so the network IO and HAProxy too. By the way, have you checked you were not using the swap? when doing load-balancing, swapping is the worst thing that could happen. Actually, the worst thing would to swap in a VM :) By the way, it seems you're running the right stud version :) Emericbr's patches has been included in mid-august. Concerning your load test, do you mean a SSL negociation per session or per connection?? You should try first with no negociation, to lower the impact on the CPU... But you're right to renogociate for each request, it provides you your total SSL handcheck capacity. Last but not least, when benching a platform, it's a bad idea to introduce random stuff. IE your "think time". How do you know how many requests your clients are generating if you introduce randomization? This might be interesting in a second step, but on first step, when benching total capacity of a platform, it's not a good idea. Well, this is my own point of view :) If you want many active connections, just create a slow backend script in PHP which holds the connection for a few seconds (let say 20) before answering "Ok". Nothing to say on your syslogs, but your rmem and wmem seems very high :) cheers On Wed, Nov 2, 2011 at 7:17 PM, Erik Torlen <erik.tor...@apicasystem.com> wrote: > Hi, > > Yeah the clients are not the problem, we are using 5 different datacenters > with 5 machines each so ~25 machines. Hardcore loadtesting :) > Btw, the loadtest are done transatlantic so that is causing latency etc. > > After some more testing yesterday we found at just what you mentioned here: > using stud with too many processes made the result much > more worse. > The perfect setup turned out to be stud with n=3 and haproxy nbproc=1. > Increasing n with n=4,5,6.. made the result much worse. > > When I got these results I used stud with n=6 which caused a lot of response > time problems. However, I don't see these response time > now when running with n=3 in haproxy logs. So how could stud with n=6 affect > the response time on the "backend" in haproxy logs? > > We are currently using the latest version of stud from github, > bumptech-stud-0.2-76-g8012fe3. Is the "emericbr patches" merge > in there or is that a fork? > > The loadtest client is doing a renegotiation for every connection. The > scenario is containing 3 small images. > Each connection is making 3 request times 10 with 3-7s waittime between each > request. > This is to maintain the connection as long as possible and get many active > connections. (We also have scenarios doing a lot of conns/s etc). > > Yeah, Aloha would have been cool to test. But this is not for us, this is for > a customer :) > > These are my main sysctl values which gave me visible performance improvement: > net.ipv4.tcp_max_syn_backlog=262144 > net.ipv4.tcp_syncookies=0 > net.ipv4.tcp_tw_reuse=1 > net.ipv4.tcp_no_metrics_save=1 > net.core.somaxconn=262144 > > net.ipv4.ip_local_port_range=1024 65536 > net.ipv4.tcp_tw_recycle=1 > > These are some more I have tried with but it did not gave me so much > improvement: > #net.ipv4.tcp_rmem=4096 87380 16777216 > #net.ipv4.tcp_wmem=4096 65536 16777216 > #net.ipv4.tcp_fin_timeout = 3 > #net.ipv4.tcp_max_orphans = 262144 > #net.ipv4.tcp_synack_retries = 2 > #net.ipv4.tcp_syn_retries = 2 > > #net.core.rmem_max=16777216 > #net.core.wmem_max=16777216 > #net.core.netdev_max_backlog = 262144 > > /E > > > -----Original Message----- > From: Baptiste [mailto:bed...@gmail.com] > Sent: den 1 november 2011 16:08 > To: Erik Torlen > Cc: haproxy@formilux.org > Subject: Re: Haproxy timing issues > > Hi, > > First question: are you sure you're reaching the limit of > haproxy/varnish and not the limit of your client? > Mainly concerning the increasing response time. > > How many CPUs do you have in your VM? Starting too much stud proccess > could be counter-productive. > I doubt doing CPU affinity in a VM improves something :) > > Concerning the logs, the time we can see on your client side are very > high! Too high :) > 3/4s for HAProxy to get the full request. > > How are you running stud? > Which options? Are you using the one with emericbr patches? > Are you using requesting using the same SSL Session ID or do you > renegotiate a new one for each connection? > > Have you checked your network statisitics, on both client and server side? > netstat -in and netstat -s > Is there a lot of drops, retransmission, congestion, etc... > > On your last log line, we can see that HAProxy took 22s to establish a > TCP connection to your Varnish... > > Can you share your stud, haproxy, and varnish configuration, the > version of each software, the startup parameters for Varnish. > What kind of tool do you use on your client to run your load test? > What sysctl have you already tunned? > > > Unfortunately, the Aloha does not run on Amazon :) > > > cheers, > > > On Tue, Nov 1, 2011 at 9:16 PM, Erik Torlen <erik.tor...@apicasystem.com> > wrote: >> Hi, >> >> I am currently (and have been from time to time the last weeks) doing some >> heavy loadtesting against haproxy with stud in front of it handling the ssl. >> >> My loadtest has been focused on loadtesting SSL traffic through stud against >> haproxy on amazon ec2. >> >> Our current problem is that we cannot get more then ~30k active connections >> (~150 conns/s) until we starting to see increased response time (>10-60s) on >> the >> client side. Running with 38k connections now and seeing much higher >> response time. >> >> The setup is: >> 1 instance running haproxy + stud >> 2 instances running varnish server 3 cached images >> >> Varnish has 100% cache hit ratio so nothing goes to the backend. >> >> We have tried using m1.xlarge and the c1.xlarge. The m1.xlarge uses almost >> 100% cpu when doing the loadtests while c1.xlarge has a lot of resources >> left (stud using a few percent per process) and haproxy ~60-70%cpu. >> The only difference is that c1.xlarge gives quite better response time >> before the actual problem happens where resp times are increasing. >> >> Haproxy is running with nbproc=1 >> Stud is running with n=6 and shared session cache. (Tried it with n=3 as well >> >> From the logging in haproxy I could see that the time it takes to establish >> a connection against the backend and receive the data: >> >> Haproxy.log >> Nov 1 18:40:35 127.0.0.1 haproxy[18511]: x.x.x.x:54113 >> [01/Nov/2011:18:39:40.273] varnish varnish/varnish1 4519/0/73/50215/54809 >> 200 2715 - - ---- 238/236/4/5/0 0/0 "GET >> /assets/images/icons/elite_logo_beta.png HTTP/1.1" >> Nov 1 18:40:35 127.0.0.1 haproxy[18511]: x.x.x.x:55635 >> [01/Nov/2011:18:39:41.547] varnish varnish/varnish1 3245/0/81/50207/53535 >> 200 1512 - - ---- 238/236/3/4/0 0/0 "GET /assets/images/icons/favicon.ico >> HTTP/1.1" >> ... >> Nov 1 18:40:44 127.0.0.1 haproxy[18511]: x.x.x.x:34453 >> [01/Nov/2011:18:39:25.330] varnish varnish/varnish1 3082/0/225/32661/79559 >> 200 1512 - - ---- 234/232/1/2/0 0/0 "GET /assets/images/icons/favicon.ico >> HTTP/1.1" >> Nov 1 18:40:44 127.0.0.1 haproxy[18511]: x.x.x.x:53731 >> [01/Nov/2011:18:39:25.036] varnish varnish/varnish1 3377/0/216/32669/79854 >> 200 1725 - - ---- 233/231/0/1/0 0/0 "GET >> /assets/images/create/action_btn.png HTTP/1.1" >> >> Haproxy.err (NOTE: 504 error here) >> >> Nov 1 18:40:11 127.0.0.1 haproxy[18511]: x.x.x.x:34885 >> [01/Nov/2011:18:39:07.597] varnish varnish/varnish1 4299/0/27/-1/64330 504 >> 194 - - sH-- 10916/10914/4777/2700/0 0/0 "GET >> /assets/images/icons/favicon.ico HTTP/1.1" >> Nov 1 18:40:12 127.0.0.1 haproxy[18511]: x.x.x.x:58878 >> [01/Nov/2011:18:39:12.621] varnish varnish/varnish2 314/0/55/-1/60374 504 >> 194 - - sH-- 3692/3690/3392/1623/0 0/0 "GET /assets/images/icons/favicon.ico >> HTTP/1.1" >> >> Nov 1 18:40:18 127.0.0.1 haproxy[18511]: x.x.x.x:35505 >> [01/Nov/2011:18:39:42.670] varnish varnish/varnish1 3515/0/22078/10217/35811 >> 200 1512 - - ---- 1482/1481/1238/710/1 0/0 "GET >> /assets/images/icons/favicon.ico HTTP/1.1" >> Nov 1 18:40:18 127.0.0.1 haproxy[18511]: x.x.x.x:40602 >> [01/Nov/2011:18:39:42.056] varnish varnish/varnish1 4126/0/22081/10226/36435 >> 200 1512 - - ---- 1475/1474/1231/703/1 0/0 "GET >> /assets/images/icons/favicon.ico HTTP/1.1" >> >> >> Here is the logs from running haproxy with varnish as a backend on the local >> machine: >> >> Haproxy.log >> Nov 1 20:00:52 127.0.0.1 haproxy[18953]: x.x.x.x:38552 >> [01/Nov/2011:20:00:45.157] varnish varnish/local_varnish 7513/0/0/0/7513 200 >> 1725 - - ---- 4/3/0/1/0 0/0 "GET /assets/images/create/action_btn.png >> HTTP/1.1" >> Nov 1 20:00:54 127.0.0.1 haproxy[18953]: x.x.x.x:40850 >> [01/Nov/2011:20:00:48.219] varnish varnish/local_varnish 6524/0/0/0/6524 200 >> 1725 - - ---- 2/1/0/1/0 0/0 "GET /assets/images/create/action_btn.png >> HTTP/1.1" >> >> Haproxy.err >> Nov 1 20:00:38 127.0.0.1 haproxy[18953]: x.x.x.x:39649 >> [01/Nov/2011:20:00:08.665] varnish varnish/local_varnish >> 7412/0/22090/23/29525 200 1511 - - ---- 15700/15698/267/268/1 0/0 "GET >> /assets/images/icons/favicon.ico HTTP/1.1" >> Nov 1 20:00:38 127.0.0.1 haproxy[18953]: x.x.x.x:54565 >> [01/Nov/2011:20:00:12.255] varnish varnish/local_varnish >> 3823/0/22090/23/25936 200 1511 - - ---- 15700/15698/266/267/1 0/0 "GET >> /assets/images/icons/favicon.ico HTTP/1.1" >> >> I see on all these tests that haproxy-stats is showing %1 idle but "top" >> shows that haproxy are using ~70% cpu? >> >> The "jungle" aka amazon and its internal network are causing a lot of >> latency when running varnish on external machine. The response times get >> better when running >> varnish locally (~0.5s improvement). >> But there is still very high response times in haproxy.err when running >> varnish locally? >> >> I have played around with sysctl values and found some that improved my >> performance. >> My feeling is that I need to tune some more values in order to go beyond >> this level, suggestions? >> >> Kind Regards >> Erik >> >> >
