-
Notifications
You must be signed in to change notification settings - Fork 340
Slow TCP #6
Comments
Thank you for checking this out and doing the tests! We will investigate the scenario you described. Indeed we have areas that need more work to improve performance. I tried your "Simple TCP client-server scripts" on win7 x64 and didn't see the crash. So the crash was in mongoDB test? Please let us know if you found an easier repro. |
There wasn't any crash. If a file is passed as the first argument, then it isn't executed. If no file is passed - the REPL is not started. For example, given file console.log('Hello World!'); on Win10 machine I get:
on Win7 machine:
|
What version of Win7 do you have? I'm on Windows 7 Enterprise Service Pack 1 and the hello.js example works.
|
My Computer Properties shows Windows 7 Enterprise Service Pack 1 64-bit. It currently has:
It is a company laptop and has McAfee running (can't disable it). |
@morkai , Regading crash on Windows 7, similar issue was recently reported at chakra-core/ChakraCore#212 and was fixed with chakra-core/ChakraCore#202. Just to make sure that this is same issue, could you please confirm if your machine has |
No |
@morkai , after investigating this for sometime, it seems that the slowness in node+chakracore is happening because of JIT optimization that we don't do for
On my dev machine, I was seeing approx. 68000 req/sec with node+v8. Before my changes mentioned above, node+chakracore was at 27000 req/sec. After my changes it boosted up to 44000 req/sec. I will continue my investigation in other gaps, but i think this is good starter point to catch up on TCP modules. |
@kunalspathak Can we test the delta between node-v8 and node-chakracore? |
Currently when running the test without an internet connection there are two JavaScript test failures and one cctest. The cctest only fails on Mac as far as I know. (I've only tested using Mac and Linux thus far). This commit moves the two JavaScript tests to test/internet. The details for test_inspector_socket_server.cc: [ RUN ] InspectorSocketServerTest.FailsToBindToNodejsHost make[1]: *** [cctest] Segmentation fault: 11 make: *** [test] Error 2 lldb output: [ RUN ] InspectorSocketServerTest.FailsToBindToNodejsHost Process 63058 stopped * thread #1: tid = 0x7b175, 0x00007fff96d04384 * libsystem_info.dylib`_gai_simple + 87, queue = * 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, * address=0x0) frame #0: 0x00007fff96d04384 libsystem_info.dylib`_gai_simple + 87 libsystem_info.dylib`_gai_simple: -> 0x7fff96d04384 <+87>: movw (%rdx), %ax 0x7fff96d04387 <+90>: movw %ax, -0x2a(%rbp) 0x7fff96d0438b <+94>: movq %r13, -0x38(%rbp) 0x7fff96d0438f <+98>: movq 0x18(%rbp), %rcx (lldb) bt * thread #1: tid = 0x7b175, 0x00007fff96d04384 * libsystem_info.dylib`_gai_simple + 87, queue = * 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, * address=0x0) * frame #0: 0x00007fff96d04384 libsystem_info.dylib`_gai_simple + 87 frame #1: 0x00007fff96cfe98b libsystem_info.dylib`search_addrinfo + 179 frame #2: 0x00007fff96cfafef libsystem_info.dylib`si_addrinfo + 2255 frame #3: 0x00007fff96cfa67b libsystem_info.dylib`getaddrinfo + 179 frame #4: 0x00000001017d8888 cctest`uv__getaddrinfo_work(w=0x00007fff5fbfe210) + 72 at getaddrinfo.c:102 frame #5: 0x00000001017d880e cctest`uv_getaddrinfo(loop=0x000000010287cb80, req=0x00007fff5fbfe1c8, cb=0x0000000000000000, hostname="nodejs.org", service="0", hints=0x00007fff5fbfe268) + 734 at getaddrinfo.c:192 frame #6: 0x000000010171f781 cctest`node::inspector::InspectorSocketServer::Start(this=0x00007fff5fbfe658) + 801 at inspector_socket_server.cc:398 frame #7: 0x00000001016ed590 cctest`InspectorSocketServerTest_FailsToBindToNodejsHost_Test::TestBody(this=0x0000000105001fd0) + 288 at test_inspector_socket_server.cc:593 I'm not sure about the exact cause for this but when using a standalone c program to simulate this it seems like when the ai_flags `AI_NUMERICSERV` is set, which is done in inspector_socket_server.cc line 394, the servname (the port in the FailsToBindToNodejsHost test) is expected to be a numeric port string to avoid looking it up in /etc/services. When the port is 0 as is it was before this commit the segment fault occurs but not if it is non-zero. PR-URL: nodejs/node#16255 Reviewed-By: Anna Henningsen <anna@addaleax.net> Reviewed-By: James M Snell <jasnell@gmail.com>
Currently when running the test without an internet connection there are two JavaScript test failures and one cctest. The cctest only fails on Mac as far as I know. (I've only tested using Mac and Linux thus far). This commit moves the two JavaScript tests to test/internet. The details for test_inspector_socket_server.cc: [ RUN ] InspectorSocketServerTest.FailsToBindToNodejsHost make[1]: *** [cctest] Segmentation fault: 11 make: *** [test] Error 2 lldb output: [ RUN ] InspectorSocketServerTest.FailsToBindToNodejsHost Process 63058 stopped * thread #1: tid = 0x7b175, 0x00007fff96d04384 * libsystem_info.dylib`_gai_simple + 87, queue = * 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, * address=0x0) frame #0: 0x00007fff96d04384 libsystem_info.dylib`_gai_simple + 87 libsystem_info.dylib`_gai_simple: -> 0x7fff96d04384 <+87>: movw (%rdx), %ax 0x7fff96d04387 <+90>: movw %ax, -0x2a(%rbp) 0x7fff96d0438b <+94>: movq %r13, -0x38(%rbp) 0x7fff96d0438f <+98>: movq 0x18(%rbp), %rcx (lldb) bt * thread #1: tid = 0x7b175, 0x00007fff96d04384 * libsystem_info.dylib`_gai_simple + 87, queue = * 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, * address=0x0) * frame #0: 0x00007fff96d04384 libsystem_info.dylib`_gai_simple + 87 frame #1: 0x00007fff96cfe98b libsystem_info.dylib`search_addrinfo + 179 frame #2: 0x00007fff96cfafef libsystem_info.dylib`si_addrinfo + 2255 frame #3: 0x00007fff96cfa67b libsystem_info.dylib`getaddrinfo + 179 frame #4: 0x00000001017d8888 cctest`uv__getaddrinfo_work(w=0x00007fff5fbfe210) + 72 at getaddrinfo.c:102 frame #5: 0x00000001017d880e cctest`uv_getaddrinfo(loop=0x000000010287cb80, req=0x00007fff5fbfe1c8, cb=0x0000000000000000, hostname="nodejs.org", service="0", hints=0x00007fff5fbfe268) + 734 at getaddrinfo.c:192 frame nodejs#6: 0x000000010171f781 cctest`node::inspector::InspectorSocketServer::Start(this=0x00007fff5fbfe658) + 801 at inspector_socket_server.cc:398 frame nodejs#7: 0x00000001016ed590 cctest`InspectorSocketServerTest_FailsToBindToNodejsHost_Test::TestBody(this=0x0000000105001fd0) + 288 at test_inspector_socket_server.cc:593 I'm not sure about the exact cause for this but when using a standalone c program to simulate this it seems like when the ai_flags `AI_NUMERICSERV` is set, which is done in inspector_socket_server.cc line 394, the servname (the port in the FailsToBindToNodejsHost test) is expected to be a numeric port string to avoid looking it up in /etc/services. When the port is 0 as is it was before this commit the segment fault occurs but not if it is non-zero. PR-URL: nodejs/node#16255 Reviewed-By: Anna Henningsen <anna@addaleax.net> Reviewed-By: James M Snell <jasnell@gmail.com>
So I've decided to see how node-chakracore v6.0.0-pre1 will compare to node-v8 v5.4.0 in generating reports in one of my apps (~500k documents fetched from MongoDB). On average node-chakracore took 2x more time to finish than node-v8.
Then I tried to import data from 27 ~8 MB text files. Parsing (CPU only) in ChakraCore was faster than in v8 (chakra=6.8s vs v8=5.6s), but then came time for some I/O (MongoDB) and there v8 won again (chakra=143s vs v8=50s).
Simple TCP client-server scripts show that the
net
module is 2x slower in ChakraCore (92kdata
events per second in v8 - 42k in ChakraCore).node and MongoDB are on the same machine (Windows 10 Home 64-bit). Tried to run it on a different machine (laptop) with Windows 7 Enterprise 64-bit, but node.exe ends immediately without any errors (
node -v
still showsv6.0.0-pre1
).The text was updated successfully, but these errors were encountered: