From 4cc3f682fb57a920e3cc14e99323d66e380c1ee7 Mon Sep 17 00:00:00 2001 From: Thomas Schwinge Date: Sat, 5 Nov 2011 23:19:34 +0100 Subject: open_issues/performance/io_system/binutils_ld_64ksec: Move Xen lseek stuff... open_issues/xen_lseek: ... here. --- .../performance/io_system/binutils_ld_64ksec.mdwn | 23 +++------------------- open_issues/performance/io_system/test-lseek.c | 17 ---------------- open_issues/performance/io_system/test-mach.c | 19 ------------------ open_issues/xen_lseek.mdwn | 22 +++++++++++++++++++++ open_issues/xen_lseek/test-lseek.c | 17 ++++++++++++++++ open_issues/xen_lseek/test-mach.c | 19 ++++++++++++++++++ 6 files changed, 61 insertions(+), 56 deletions(-) delete mode 100644 open_issues/performance/io_system/test-lseek.c delete mode 100644 open_issues/performance/io_system/test-mach.c create mode 100644 open_issues/xen_lseek/test-lseek.c create mode 100644 open_issues/xen_lseek/test-mach.c diff --git a/open_issues/performance/io_system/binutils_ld_64ksec.mdwn b/open_issues/performance/io_system/binutils_ld_64ksec.mdwn index d0b8ea7f..931fd0ee 100644 --- a/open_issues/performance/io_system/binutils_ld_64ksec.mdwn +++ b/open_issues/performance/io_system/binutils_ld_64ksec.mdwn @@ -27,10 +27,6 @@ extracted from cdf7c161ebd4a934c9e705d33f5247fd52975612 sources, 2010-10-24. On the idle grubber, this one repeatedly takes a few minutes wall time to complete successfully, contrary to a few seconds on a GNU/Linux system. -> On order of slowness may in fact be due to a Xen-specific issue, see -> [[xen_lseek]]. (But there are probably still one or two orders left, even -> without Xen.) - While processing the object files, there is heavy interaction with the relevant [[hurd/translator/ext2fs]] process. Running [[hurd/debugging/rpctrace]] on the testee shows that (primarily) an ever-repeating series of `io_seek` and @@ -38,19 +34,6 @@ the testee shows that (primarily) an ever-repeating series of `io_seek` and shows the equivalent thing (`_llseek`, `read`) -- but Linux' I/O system isn't as slow as the Hurd's. ---- - -IRC, freenode, #hurd, 2011-09-01: - - hum, f951 does myriads of 71->io_seek_request (32768 0) = 0 32768 - no wonder it's slow - unfortunately that's also what it does on linux, the system call is - just less costly - apparently gfortran calls io_seek for, like, every token of the - sourced file - (fgetpos actually, but that's the same) - and it is indeed about 10 times slower under Xen for some reason - -Also see testcase [[test-lseek.c]] and [[test-mach.c]] - -[[!tag open_issue_xen]] +As Samuel figured out later, this slowness may in fact be due to a Xen-specific +issue, see [[Xen_lseek]]. After the latter has been addressed, we can +re-evaluate this issue here. diff --git a/open_issues/performance/io_system/test-lseek.c b/open_issues/performance/io_system/test-lseek.c deleted file mode 100644 index 667dce66..00000000 --- a/open_issues/performance/io_system/test-lseek.c +++ /dev/null @@ -1,17 +0,0 @@ -#include -#include -#include -#include -#include -#define N 100000 -int main(void) { - int fd = open("test.c", O_RDONLY); - struct timeval tv1, tv2; - int i; - gettimeofday(&tv1, NULL); - for (i = 0; i < N; i++) - lseek(fd, 0, SEEK_CUR); - gettimeofday(&tv2, NULL); - printf("%fµs\n", (float)((tv2.tv_sec-tv1.tv_sec) * 1000000 + tv2.tv_usec - tv1.tv_usec)/N); - return 0; -} diff --git a/open_issues/performance/io_system/test-mach.c b/open_issues/performance/io_system/test-mach.c deleted file mode 100644 index 90337346..00000000 --- a/open_issues/performance/io_system/test-mach.c +++ /dev/null @@ -1,19 +0,0 @@ -#define _GNU_SOURCE -#include -#include -#include -#define N 1000000 -int main(void) { - struct timeval tv1, tv2; - int i; - task_t task; - task = mach_task_self(); - mach_port_urefs_t refs; - gettimeofday(&tv1, NULL); - for (i = 0; i < N; i++) { - mach_port_get_refs(task, task, MACH_PORT_RIGHT_RECEIVE, &refs); - } - gettimeofday(&tv2, NULL); - printf("%fµs\n", (float)((tv2.tv_sec-tv1.tv_sec) * 1000000 + tv2.tv_usec - tv1.tv_usec)/N); - return 0; -} diff --git a/open_issues/xen_lseek.mdwn b/open_issues/xen_lseek.mdwn index accc7c8f..756abf5e 100644 --- a/open_issues/xen_lseek.mdwn +++ b/open_issues/xen_lseek.mdwn @@ -10,6 +10,17 @@ License|/fdl]]."]]"""]] [[!tag open_issue_xen]] +IRC, freenode, #hurd, 2011-09-01: + + hum, f951 does myriads of 71->io_seek_request (32768 0) = 0 32768 + no wonder it's slow + unfortunately that's also what it does on linux, the system call is + just less costly + apparently gfortran calls io_seek for, like, every token of the + sourced file + (fgetpos actually, but that's the same) + and it is indeed about 10 times slower under Xen for some reason + IRC, freenode, #hurd, 2011-11-02: btw, we have a performance issue with xen @@ -33,3 +44,14 @@ IRC, freenode, #hurd, 2011-11-02: http://www.gnu.org/software/hurd/open_issues/performance/io_system/binutils_ld_64ksec.html [[performance/io_system/binutils_ld_64ksec]]. + +Also see the simple testcases [[test-lseek.c]] and [[test-mach.c]]. + +IRC, freenode, #hurd, 2011-11-05: + + [test-mach.c is] mostly as a reference for the trap overhead + 0.56µs (xen) vs 0.48µs(kvm) on test-mach + 455µs(xen) vs 16µs(kvm) on test-lseek + that might simply be an issue in the RPC mechanism, which behaves + badly with the xen memory management + yes, about 0.5ms for an lseek, that's quite high :) diff --git a/open_issues/xen_lseek/test-lseek.c b/open_issues/xen_lseek/test-lseek.c new file mode 100644 index 00000000..667dce66 --- /dev/null +++ b/open_issues/xen_lseek/test-lseek.c @@ -0,0 +1,17 @@ +#include +#include +#include +#include +#include +#define N 100000 +int main(void) { + int fd = open("test.c", O_RDONLY); + struct timeval tv1, tv2; + int i; + gettimeofday(&tv1, NULL); + for (i = 0; i < N; i++) + lseek(fd, 0, SEEK_CUR); + gettimeofday(&tv2, NULL); + printf("%fµs\n", (float)((tv2.tv_sec-tv1.tv_sec) * 1000000 + tv2.tv_usec - tv1.tv_usec)/N); + return 0; +} diff --git a/open_issues/xen_lseek/test-mach.c b/open_issues/xen_lseek/test-mach.c new file mode 100644 index 00000000..90337346 --- /dev/null +++ b/open_issues/xen_lseek/test-mach.c @@ -0,0 +1,19 @@ +#define _GNU_SOURCE +#include +#include +#include +#define N 1000000 +int main(void) { + struct timeval tv1, tv2; + int i; + task_t task; + task = mach_task_self(); + mach_port_urefs_t refs; + gettimeofday(&tv1, NULL); + for (i = 0; i < N; i++) { + mach_port_get_refs(task, task, MACH_PORT_RIGHT_RECEIVE, &refs); + } + gettimeofday(&tv2, NULL); + printf("%fµs\n", (float)((tv2.tv_sec-tv1.tv_sec) * 1000000 + tv2.tv_usec - tv1.tv_usec)/N); + return 0; +} -- cgit v1.2.3