java - "Lost" UDP packets (JBoss + DatagramSocket) -

-

i develop part of jboss+ejb based enterprise application. module needs process huge amount of incoming udp packets. i've done load testing , looks in case of sending packets 11ms interval fine, in case of 10ms interval packets lost. it's rather strange in opinion, done 10/11ms interval load tests comparison several times , same result (10 ms - "lost" packets, 11ms - everything's fine).

if wrong synchronization, i'd expect visible in case of 11ms tests (at least 1 packet lost, or @ least 1 wrong counter value). if not because of synchronization, maybe datagramsocket through receive packets doesn't work expected.

i found receive buffer size (so_rcvbuf) has default 57344 value (probably it's underlying io network buffers dependent). suspect, maybe when buffer goes full, new incoming udp datagrams rejected. tried set value higher, noticed if exaggerate, buffer returns default size. if it's underlying layer dependent how can find out maximum buffer size os/network card jboss level?

is possible caused receive buffer size, or maybe 57344 value big enough handle cases? have experience such issues?

there no timeout set on datagramsocket. udp datagrams contains 70 bytes of data (value without datagram header included).

[edited] have use udp because receive cisco netflow data - protocol used network devices send traffic statistics. also, have no influence on sent bytes format (e.g. cannot add counters packets , on). not expected packets processed (some datagrams may lost), i'd expect process of packets. during 10ms interval tests, 30% of packets lost.

it not possible slow processing causes issue. singleton component holds reference datagramsocket calling receive method in loop. when packet received, passed queue, , processed picked pool stateless component. "facade" singleton responsible receiving packets , passing on processing (it not wait processing complete event).

thanks in advance, piotr

udp inherently unreliable.

datagrams can thrown away @ point between sender , receiver, within receiver @ level below code. setting recv buffer larger size networking code within machine buffer more datagrams should expect datagrams lost anyway.

if recv logic takes long (i.e. longer takes new datagram arrive) you'll behind , you'll miss datagrams eventually. can make sure recv code runs fast possible, perhaps move inbound datagram queue , process 'later' or on thread move problem being 1 have queue keeps growing.

[re edit...] , what's processing queue , how locking work between producer , consumers? change code recv logic increments count , discards data , loops around , see if you're losing fewer datagrams; either way, udp unreliable, have datagrams discarded , should expect , deal it. worrying means you're focusing on wrong problem; make use of data , assume wont of , program work if network gets congested , of datagrams discarded.

in summary, that's how udp.


Comments

Post a Comment

Popular posts from this blog

linux - Mailx and Gmail nss config dir -

-
i trying send mail linux command line using mailx command. can send local domain no problem want set mail send gmail account receive mail sent gmail account. after configuring mail.rc so: account gmail { set smtp=smtps://smtp.gmail.com:587 set smtp-auth=login set smtp-auth-user=username@gmail.com set smtp-auth-password=password set ssl-verify=ignore } i error: resolving host smtp.gmail.com . . . done. connecting 74.125.25.109 . . . connected. missing "nss-config-dir" variable. "/home/username/dead.letter" 11/354 . . . message not sent. after looking "nss-config-dir" here , located certn.db , keyn.db files , added mail.rc so: account gmail { set smtp=smtps://smtp.gmail.com:587 set smtp-auth=login set smtp-auth-user=username@gmail.com set smtp-auth-password=password set ssl-verify=ignore set nss-config-dir=/home/user/.mozilla/firefox/location.default } now when try send mail using command: echo "sent gmail account" | mailx -v -a
Read more

c# - Is it possible to remove an existing registration from Autofac container builder? -

-
something along lines: builder.registertype<mytype>().as<itype>(); builder.registertype<mytype2>().as<itype>(); builder.deregistertype<mytype>().as<itype>() var container = builder.build(); var types = container.resolve<ienumerable<itype>>(); assert.istrue(types.count == 1); assert.istrue(types[0].gettype == typeof(mytype2)); scenario: go through bunch of assemblies , go register types want make sure have 1 implementation of given type. need before create container. track on own nice if autofac me bit. this cannot done directly using containerbuilder , unless start on new one. mind you, having first built container should able construct new container filtering away unwanted types , reusing registrations first container. this: ... var container = builder.build(); builder = new containerbuilder(); var components = container.componentregistry.registrations .where(cr => cr.activator.limittype != typ
Read more

php - Mysql PK and FK char(36) vs int(10) -

-
i have read mysql can handle indexes , searches better if tables keys integers rather char(36) uuid's. is worth converting int(10)? note: use 1 server , have have plans use multiple databases concurrently therefore removing 1 of big reasons using uuid. our tables innodb if makes difference. thanks in advance. usually rdbms can handle integer keys pk more efficient, other datatypes. reason how build index column, yes: long dont require string (or other typed) keys, should use integers. however: char(36) , int(10) far away being equal, because int(10) much smaller, char(36). dont know, if require many different keys, should keep in mind. update, complete last paragraph: int(10) 32-bit, char(36) 36 - ^byte^ (= 288-bit). not mean, int consumes less space, means, char(36) has 4 times more different keys.
Read more