Skip to contentSkip to conversation

Recent slowdown in single object builds

DISCUSSION Jens Axboe 2 participants 5 messages Nov 15, 2019 – Nov 18, 2019

lore
Share & tools
lore

Jens Axboe

author
#0 lore raw

Hi,

I've noticed that current -git is a lot slower at doing single object builds than earlier kernels. Here's an example, building the exact same file on 5.2 and -git:

$ time make fs/io_uring.o
real    0m5.953s
user    0m5.402s
sys     0m0.649s

vs 5.2 based (with all the backports, identical file):

$ time make fs/io_uring.o
real    0m3.218s
user    0m2.968s
sys     0m0.520s

Any idea what's going on here? It's almost twice as slow, which is problematic...

Signature · 1 line
Jens Axboe
#1 lore raw
Quoted 14 lines · Jens Axboe

Hi,

I've noticed that current -git is a lot slower at doing single object builds than earlier kernels. Here's an example, building the exact same file on 5.2 and -git:

$ time make fs/io_uring.o
real    0m5.953s
user    0m5.402s
sys     0m0.649s

vs 5.2 based (with all the backports, identical file):

Quoted 7 lines
$ time make fs/io_uring.o
real    0m3.218s
user    0m2.968s
sys     0m0.520s

Any idea what's going on here? It's almost twice as slow, which is problematic...

This is necessary cost
to do single builds
(394053f4a4b3e3eeeaa67b67fc886a9a75bd9e4d)
but, it is much better in linux-next.

Signature · 2 lines
Best Regards
Masahiro Yamada
author
#2 lore raw
Quoted 25 lines · Masahiro Yamada
Quoted 14 lines · Jens Axboe

Hi,

I've noticed that current -git is a lot slower at doing single object builds than earlier kernels. Here's an example, building the exact same file on 5.2 and -git:

$ time make fs/io_uring.o
real    0m5.953s
user    0m5.402s
sys     0m0.649s

vs 5.2 based (with all the backports, identical file):

Quoted 7 lines
$ time make fs/io_uring.o
real    0m3.218s
user    0m2.968s
sys     0m0.520s

Any idea what's going on here? It's almost twice as slow, which is problematic...

Quoted 4 lines

This is necessary cost
to do single builds
(394053f4a4b3e3eeeaa67b67fc886a9a75bd9e4d)
but, it is much better in linux-next.

Very sad that it's now twice as slow as before, that's a real problem. This isn't a marginal slowdown.

I've never had issues with single object builds before, and in fact it's often useful to build stuff that's disabled as a single object. Where's the report that led to this commit being necessary?

Signature · 1 line
Jens Axboe
#3 lore raw
Quoted 17 lines · Jens Axboe
Quoted 15 lines · Masahiro Yamada
Quoted 14 lines · Jens Axboe

Hi,

I've noticed that current -git is a lot slower at doing single object builds than earlier kernels. Here's an example, building the exact same file on 5.2 and -git:

$ time make fs/io_uring.o
real    0m5.953s
user    0m5.402s
sys     0m0.649s

vs 5.2 based (with all the backports, identical file):

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
$ time make fs/io_uring.o
real    0m3.218s
user    0m2.968s
sys     0m0.520s

Any idea what's going on here? It's almost twice as slow, which is problematic...

BTW, I am using a cheap machine based on Intel Core i7-6700K. I can build fs/io_uring.o (based on x86_64_defconfig) much faster.

The slowdown (1.990 sec -> 2.861 sec) is not terrible for me.

masahiro@pug:~/ref/linux$ git describe
v5.2
masahiro@pug:~/ref/linux$ touch  fs/io_uring.c; time  make  fs/io_uring.o
  CALL    scripts/checksyscalls.sh
  CALL    scripts/atomic/check-atomics.sh
  DESCEND  objtool
  CC      fs/io_uring.o

real 0m1.990s
user 0m1.723s
sys 0m0.427s

masahiro@pug:~/ref/linux-next$ git describe
next-20191115
masahiro@pug:~/ref/linux-next$ touch   fs/io_uring.c; time  make  fs/io_uring.o
  CALL    scripts/checksyscalls.sh
  CALL    scripts/atomic/check-atomics.sh
  DESCEND  objtool
  CC      fs/io_uring.o

real 0m2.861s
user 0m2.373s
sys 0m0.642s

Snippet from my /proc/cpuinfo

processor : 0
vendor_id : GenuineIntel
cpu family : 6
model : 94
model name : Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz
stepping : 3
microcode : 0xd4
cpu MHz : 800.023
cache size : 8192 KB

Quoted 17 lines
Quoted 6 lines

This is necessary cost
to do single builds
(394053f4a4b3e3eeeaa67b67fc886a9a75bd9e4d)
but, it is much better in linux-next.

Very sad that it's now twice as slow as before, that's a real problem. This isn't a marginal slowdown.

I've never had issues with single object builds before, and in fact it's often useful to build stuff that's disabled as a single object. Where's the report that led to this commit being necessary?

--
Jens Axboe

Signature · 2 lines
Best Regards
Masahiro Yamada
author
#4 lore raw
Quoted 27 lines · Masahiro Yamada
Quoted 17 lines · Jens Axboe
Quoted 15 lines · Masahiro Yamada
Quoted 14 lines · Jens Axboe

Hi,

I've noticed that current -git is a lot slower at doing single object builds than earlier kernels. Here's an example, building the exact same file on 5.2 and -git:

$ time make fs/io_uring.o
real    0m5.953s
user    0m5.402s
sys     0m0.649s

vs 5.2 based (with all the backports, identical file):

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
$ time make fs/io_uring.o
real    0m3.218s
user    0m2.968s
sys     0m0.520s

Any idea what's going on here? It's almost twice as slow, which is problematic...

Quoted 5 lines

BTW, I am using a cheap machine based on Intel Core i7-6700K. I can build fs/io_uring.o (based on x86_64_defconfig) much faster.

The slowdown (1.990 sec -> 2.861 sec) is not terrible for me.

For those of us that do single object builds all the time the slowdown is pretty damn terrible. Granted that your 44% slowdown isn't as bad as my 85% slowdown, but it's still a MASSIVE slowdown.

Signature · 1 line
Jens Axboe