Skip to contentSkip to conversation

Improve performance for native/cross/crosssdk's cleansstate

PATCH Superseded Robert Yang 2 participants 5 messages Nov 22, 2018

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Robert Yang

author
#0 lore raw

cover Improve performance for native/cross/crosssdk's cleansstate

Hi RP and Ross,

These 2 patches can improve performance a lot for native/cross/crosssdk's cleansstate:

This test is based on more than 600,000 sstate files
* Without disk caches

  # echo 3 >/proc/sys/vm/drop_caches
  $ bitbake -p
  $ time bitbake quilt-native -ccleansstate
  - Before:
    real    4m53.815s
    user    0m0.820s
    sys     0m0.128s

- After

    real    0m58.483s
    user    0m0.744s
    sys     0m0.108s

Saved 293 - 58 = 235s (80% improvement)

* With disk caches (26.6s -> 5.8s, 78% improvement)

  $ bitbake -p
  $ time bitbake quilt-native -ccleansstate
  - Before:
    real    0m26.682s
    user    0m0.748s
    sys     0m0.088s

- After

    real    0m5.846s
    user    0m0.716s
    sys     0m0.116s

Saved 26.6 - 5.8 = 20.8s (78% improvement)

======
I have another idea is put sstate files into the directory named by the recipe, e.g.:

sstate-cache/quilt-native/sstate:quilt-native::0.65:r0::3:773341447192feb6c1519cace0e6be35_populate_lic.tgz sstate-cache/quilt-native/sstate:quilt-native::0.65:r0::3:773341447192feb6c1519cace0e6be35_populate_lic.tgz.siginfo sstate-cache/quilt-native/sstate:quilt-native::0.65:r0::3:ab4a82a07b10926bb42d2f37242413c7_patch.tgz.siginfo sstate-cache/quilt-native/sstate:quilt-native::0.65:r0::3:e8a4c952a66942653e36f289eaf68ca5_fetch.tgz.siginfo sstate-cache/quilt-native/sstate:quilt-native::0.65:r0::3:fd35240c437cefa66e24cfe65d3fb335_unpack.tgz.siginfo

This can make it easier to manage sstate files, and can reduce the load of glob.glob(), currently, the remove command is something like:

$ rm -f glob.glob(sstate-cache/*/sstate:foo.*bar.*)

If we put them into recipe named directory, then we can run:

$ rm -f glob.glob(sstate-cache/<recipe>/sstate:foo.*bar.*)

Which should be faster, what's your opinion, please ?

// Robert

The following changes since commit fc5418e7bbdecfb27bafe595084e0fd0f991a388:

meta/icecc.bbclass: Update system blacklists (2018-11-21 11:48:18 +0000)

are available in the git repository at:

git://git.openembedded.org/openembedded-core-contrib rbt/sstate
http://cgit.openembedded.org/openembedded-core-contrib/log/?h=rbt/sstate

Robert Yang (2):
nopackages.bbclass: improve performance for cleansstate sstate.bbclass: set SSTATE_EXTRAPATHWILDCARD explicitly

 meta/classes/nopackages.bbclass | 35 +++++++++++++++++++++++------------
 meta/classes/sstate.bbclass     |  2 +-
 2 files changed, 24 insertions(+), 13 deletions(-)
Signature · 1 line
2.7.4
author
#1 lore raw

PATCH 1/2 nopackages.bbclass: improve performance for cleansstate

Leave the task in SSTATETASKS doesn't make sense when it is removed, so also remove it from SSTATETASKS.

This can improve the performance a lot for "bitbake <recipe-native/cross/crosssdk> -ccleansstate" when there are a lot of sstate files.

For example:
* Before

  $ bitbake quilt-native -ccleansstate
  - Check log.do_cleansstate:
  Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-linux:0.65:r0:x86_64:3:*_package.tgz*
  Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-linux:0.65:r0:x86_64:3:*_package_qa.tgz*
  Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-linux:0.65:r0:x86_64:3:*_package_write_rpm.tgz*
  Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-linux:0.65:r0:x86_64:3:*_packagedata.tgz*
  Removing /sstate-cache/*/sstate:quilt-native::0.65:r0::3:*_populate_lic.tgz*
  Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-linux:0.65:r0:x86_64:3:*_populate_sysroot.tgz*

There are no package tasks for quilt-native, so the first 4 lines doesn't make any sense, but the glob pattern "sstate-cache/*/*" is very time consuming when there are no disk caches. E.g., I have more than 600,000 sstate files:
- Without disk caches

  # echo 3 >/proc/sys/vm/drop_caches
  $ time python3 -c 'import glob; glob.glob("/sstate-cache/*/*/sstate:quilt-native:x86_64-linux:0.65:r0:x86_64:3:*_package.tgz*")'
    real    4m32.583s
    user    0m5.768s
    sys     0m12.892s

- With disk caches (e.g., run it in the second time)

  $ time python3 -c 'import glob; glob.glob("/sstate-cache/*/*/sstate:quilt-native:x86_64-linux:0.65:r0:x86_64:3:*_package.tgz*")'
    real    0m5.128s
    user    0m2.772s
    sys     0m2.308s

So the 4 removing *package* commands cost more than 20s or 272s in theory.

* After

  $ bitbake quilt-native -ccleansstate
  - Check log.do_cleansstate:
  Removing /sstate-cache/*/sstate:quilt-native::0.65:r0::3:*_populate_lic.tgz*
  Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-linux:0.65:r0:x86_64:3:*_populate_sysroot.tgz*

Now the package tasks are gone when cleanssate for native recipes, which saved 20s or 272s in theory.

Signed-off-by Robert Yang <[email protected]>
---
 meta/classes/nopackages.bbclass | 35 +++++++++++++++++++++++------------
 1 file changed, 23 insertions(+), 12 deletions(-)
meta/classes/nopackages.bbclass +23 -12
index 559f507..66e4936 100644
@@ -1,12 +1,23 @@
-deltask do_package
-deltask do_package_write_rpm
-deltask do_package_write_ipk
-deltask do_package_write_deb
-deltask do_package_qa
-deltask do_packagedata
-deltask do_package_setscene
-deltask do_package_write_rpm_setscene
-deltask do_package_write_ipk_setscene
-deltask do_package_write_deb_setscene
-deltask do_package_qa_setscene
-deltask do_packagedata_setscene
+NO_PACKAGES_TASKS = " \
+    do_package \
+    do_package_write_rpm \
+    do_package_write_ipk \
+    do_package_write_deb \
+    do_package_qa \
+    do_packagedata \
+"
+
+python() {
+    sstatetasks = (d.getVar('SSTATETASKS') or '').split()
+    for task in d.getVar('NO_PACKAGES_TASKS').split():
+        setscene_task = '%s_setscene' % task
+        tasks = (task, setscene_task)
+        for t in tasks:
+            bb.build.deltask(t, d)
+
+        if task in sstatetasks:
+            bb.debug(2, 'Removing sstate task %s from SSTATETASKS' % task)
+            sstatetasks.remove(task)
+    d.setVar('SSTATETASKS', ' '.join(sstatetasks))
+
+}
Signature · 1 line
2.7.4
#2 lore raw

Re: [PATCH 1/2] nopackages.bbclass: improve performance for cleansstate

Quoted 108 lines · Robert Yang

Leave the task in SSTATETASKS doesn't make sense when it is removed, so also
remove it from SSTATETASKS.

This can improve the performance a lot for "bitbake <recipe- native/cross/crosssdk>
-ccleansstate" when there are a lot of sstate files.

For example:
* Before

  $ bitbake quilt-native -ccleansstate
  - Check log.do_cleansstate:
  Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_package.tgz*
  Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_package_qa.tgz*
  Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_package_write_rpm.tgz*
  Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_packagedata.tgz*
  Removing /sstate-cache/*/sstate:quilt-
native::0.65:r0::3:*_populate_lic.tgz*
  Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_populate_sysroot.tgz*

There are no package tasks for quilt-native, so the first 4 lines
doesn't
make any sense, but the glob pattern "sstate-cache/*/*" is very
time
consuming when there are no disk caches. E.g., I have more than
600,000
sstate files:
- Without disk caches

  # echo 3 >/proc/sys/vm/drop_caches
  $ time python3 -c 'import glob; glob.glob("/sstate-
cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_package.tgz*")'
    real    4m32.583s
    user    0m5.768s
    sys     0m12.892s

- With disk caches (e.g., run it in the second time)

  $ time python3 -c 'import glob; glob.glob("/sstate-
cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_package.tgz*")'
    real    0m5.128s
    user    0m2.772s
    sys     0m2.308s

So the 4 removing *package* commands cost more than 20s or 272s in
theory.

* After

  $ bitbake quilt-native -ccleansstate
  - Check log.do_cleansstate:
  Removing /sstate-cache/*/sstate:quilt-
native::0.65:r0::3:*_populate_lic.tgz*
  Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_populate_sysroot.tgz*

Now the package tasks are gone when cleanssate for native recipes,
which
saved 20s or 272s in theory.

Signed-off-by Robert Yang <[email protected]>
---
 meta/classes/nopackages.bbclass | 35 +++++++++++++++++++++++------
------
 1 file changed, 23 insertions(+), 12 deletions(-)
meta/classes/nopackages.bbclass +19 -12
index 559f507..66e4936 100644
@@ -1,12 +1,23 @@
-deltask do_package
-deltask do_package_write_rpm
-deltask do_package_write_ipk
-deltask do_package_write_deb
-deltask do_package_qa
-deltask do_packagedata
-deltask do_package_setscene
-deltask do_package_write_rpm_setscene
-deltask do_package_write_ipk_setscene
-deltask do_package_write_deb_setscene
-deltask do_package_qa_setscene
-deltask do_packagedata_setscene
+NO_PACKAGES_TASKS = " \
+    do_package \
+    do_package_write_rpm \
+    do_package_write_ipk \
+    do_package_write_deb \
+    do_package_qa \
+    do_packagedata \
+"
+
+python() {
+    sstatetasks = (d.getVar('SSTATETASKS') or '').split()
+    for task in d.getVar('NO_PACKAGES_TASKS').split():
+        setscene_task = '%s_setscene' % task
+        tasks = (task, setscene_task)
+        for t in tasks:
+            bb.build.deltask(t, d)
+
+        if task in sstatetasks:
+            bb.debug(2, 'Removing sstate task %s from SSTATETASKS' %

task)

meta/classes/nopackages.bbclass line 23

Quoted 4 lines
meta/classes/nopackages.bbclass +4 -0
+            sstatetasks.remove(task)
+    d.setVar('SSTATETASKS', ' '.join(sstatetasks))
+
+}

I really don't want to turn more comparatively readable classes into a mess of anonymous python if we can help it.

Can we not just teach cleansstate to check if a task exists before trying to delete the sstate? It should be possible with d.getVarFlag(taskname, "task")?

Cheers,

Richard

author
#3 lore raw

Re: [PATCH 1/2] nopackages.bbclass: improve performance for cleansstate

Sent from mobile phone

Quoted 119 lines · 在 2018年11月22日，20:15，Richard Purdie <[email protected]> 写道：
Quoted 110 lines · Robert Yang

Leave the task in SSTATETASKS doesn't make sense when it is removed, so also
remove it from SSTATETASKS.

This can improve the performance a lot for "bitbake <recipe- native/cross/crosssdk>
-ccleansstate" when there are a lot of sstate files.

For example:
* Before

 $ bitbake quilt-native -ccleansstate
 - Check log.do_cleansstate:
 Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_package.tgz*
 Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_package_qa.tgz*
 Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_package_write_rpm.tgz*
 Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_packagedata.tgz*
 Removing /sstate-cache/*/sstate:quilt-
native::0.65:r0::3:*_populate_lic.tgz*
 Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_populate_sysroot.tgz*

There are no package tasks for quilt-native, so the first 4 lines
doesn't
make any sense, but the glob pattern "sstate-cache/*/*" is very
time
consuming when there are no disk caches. E.g., I have more than
600,000
sstate files:
- Without disk caches

 # echo 3 >/proc/sys/vm/drop_caches
 $ time python3 -c 'import glob; glob.glob("/sstate-
cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_package.tgz*")'
   real    4m32.583s
   user    0m5.768s
   sys     0m12.892s

- With disk caches (e.g., run it in the second time)

 $ time python3 -c 'import glob; glob.glob("/sstate-
cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_package.tgz*")'
   real    0m5.128s
   user    0m2.772s
   sys     0m2.308s

So the 4 removing *package* commands cost more than 20s or 272s in
theory.

* After

 $ bitbake quilt-native -ccleansstate
 - Check log.do_cleansstate:
 Removing /sstate-cache/*/sstate:quilt-
native::0.65:r0::3:*_populate_lic.tgz*
 Removing /sstate-cache/*/*/sstate:quilt-native:x86_64-
linux:0.65:r0:x86_64:3:*_populate_sysroot.tgz*

Now the package tasks are gone when cleanssate for native recipes,
which
saved 20s or 272s in theory.

Signed-off-by Robert Yang <[email protected]>

---
meta/classes/nopackages.bbclass | 35 +++++++++++++++++++++++------
------

1 file changed, 23 insertions(+), 12 deletions(-)
meta/classes/nopackages.bbclass +19 -12
index 559f507..66e4936 100644
@@ -1,12 +1,23 @@
-deltask do_package
-deltask do_package_write_rpm
-deltask do_package_write_ipk
-deltask do_package_write_deb
-deltask do_package_qa
-deltask do_packagedata
-deltask do_package_setscene
-deltask do_package_write_rpm_setscene
-deltask do_package_write_ipk_setscene
-deltask do_package_write_deb_setscene
-deltask do_package_qa_setscene
-deltask do_packagedata_setscene
+NO_PACKAGES_TASKS = " \
+    do_package \
+    do_package_write_rpm \
+    do_package_write_ipk \
+    do_package_write_deb \
+    do_package_qa \
+    do_packagedata \
+"
+
+python() {
+    sstatetasks = (d.getVar('SSTATETASKS') or '').split()
+    for task in d.getVar('NO_PACKAGES_TASKS').split():
+        setscene_task = '%s_setscene' % task
+        tasks = (task, setscene_task)
+        for t in tasks:
+            bb.build.deltask(t, d)
+
+        if task in sstatetasks:
+            bb.debug(2, 'Removing sstate task %s from SSTATETASKS' %

task)

meta/classes/nopackages.bbclass line 23

Quoted 4 lines
meta/classes/nopackages.bbclass +4 -0
+            sstatetasks.remove(task)
+    d.setVar('SSTATETASKS', ' '.join(sstatetasks))
+
+}

I really don't want to turn more comparatively readable classes into a mess of anonymous python if we can help it.

Quoted 3 lines

Can we not just teach cleansstate to check if a task exists before trying to delete the sstate? It should be possible with d.getVarFlag(taskname, "task")?

Thanks, yes, this is my first thought, but I only knew __BBTASKS which is an internal var of bitbake, I will use d.getVarFlag(taskname, "task") to send a v2.

// Robert

Quoted 5 lines

Cheers,

Richard

author
#4 lore raw

PATCH 2/2 sstate.bbclass: set SSTATE_EXTRAPATHWILDCARD explicitly

In reply to Robert Yang

The glob.glob("/sstate/*/*/") is very time consuming, set SSTATE_EXTRAPATHWILDCARD explicity to avoid that. This can save a lot of time when there are many sstate files.

For example, I have more than 600,000 sstate files:
* Before
- Without disk caches

  $ time python3 -c 'import glob; glob.glob("/sstate-cache/*/*/sstate:quilt-native:x86_64-linux:0.65:r0:x86_64:3:*_populate_sysroot.tgz*")'
    real    4m32.583s
    user    0m5.768s
    sys     0m12.892s

- With disk caches

  $ time python3 -c 'import glob; glob.glob("/sstate-cache/*/*/sstate:quilt-native:x86_64-linux:0.65:r0:x86_64:3:*_populate_sysroot.tgz*")'
  real    0m4.111s
  user    0m2.348s
  sys     0m1.756s

* After

  $ time python3 -c 'import glob; glob.glob("/sstate-cache.bak/universal/*/sstate:quilt-native:x86_64-linux:0.65:r0:x86_64:3:*_populate_sysroot.tgz*")'
  - Without disk caches:
  real    0m7.928s
  user    0m0.172s
  sys     0m0.124s

- With disk caches:

  real    0m0.131s
  user    0m0.088s
  sys     0m0.044s

We can see that it saves about 3.8s with disk caches, and saves about 264s without disk caches.

Signed-off-by Robert Yang <[email protected]>
---
 meta/classes/sstate.bbclass | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
meta/classes/sstate.bbclass +1 -1
index efb0096..94fde6d 100644
@@ -101,7 +101,7 @@ python () {
     if bb.data.inherits_class('native', d) or bb.data.inherits_class('crosssdk', d) or bb.data.inherits_class('cross', d):
         d.setVar('SSTATE_EXTRAPATH', "${NATIVELSBSTRING}/")
         d.setVar('BB_HASHFILENAME', "True ${SSTATE_PKGSPEC} ${SSTATE_SWSPEC}")
-        d.setVar('SSTATE_EXTRAPATHWILDCARD', "*/")
+        d.setVar('SSTATE_EXTRAPATHWILDCARD', "${NATIVELSBSTRING}/")
 
     unique_tasks = sorted(set((d.getVar('SSTATETASKS') or "").split()))
     d.setVar('SSTATETASKS', " ".join(unique_tasks))
Signature · 1 line
2.7.4