Proposal to set dalvik-vm optimizing all code

hunderteins · Apr 27, 2012

Hello,

I played with the dalvik.vm.dexopt-flags in /system/build.prop. Basically you control the class-loading and dalvik-vm there.

streakowner proposed in http://xdaforums.com/showpost.php?p=18111740&postcount=644 the setting v=n,o=v. But disabled this with #, in DSC v1.0 the setting is active.
That means: do not verify and optimize mode verified. Found that info in dexopt/OptMain.c

I did some performance measurement with smartbench 2011 and got the value 553 in the first round, and 465 in the second round. That is very bad for an one ghz powered device. Good would > 1000. /data/dalvik-cache takes 74 MiB with this setting.

At the moment I'm running with

dalvik.vm.dexopt-flags=v=n,o=a,u=y

That means: do not verify, optimize all classes and optimize for uniprocessor.

First I was not impressed, because rebooting with a cleaned dalvik-cache took more than double the time span of the o=v setting. logcat is full off lines telling me what can't be optimized and how much time the load took. But after booting I got the impression the system went faster. With smartbench 2011 I get the value 1036 in the first and 1110 in the second round. /data/dalvik-cache takes 74 MiB with this setting.

Can someone confirm my finding?

I don't say you will double the speed of the streak with this setting. But I think you will notice the change. The trade off is the time for the first boot. But I think that is worth is.
So I propose to use the above setting in all gb-based roms for the streak 5.

Have a nice weekend,
hunderteins

_2bad4u_ · Apr 27, 2012

Tested on miui and it's not working i get into a boot loop.

_n0p_ · Apr 27, 2012

getprop | grep dexopt
[dalvik.vm.dexopt-flags]: [v=n,o=a,u=y]

Well, same ~950 in Quadrant (CPU/mem only)

998400

--
Sergei.

hunderteins · Apr 27, 2012

_2bad4u_ said:
Tested on miui and it's not working i get into a boot loop.

Ouch! Sorry to hear that. I hope you didn't get in serious trouble.

_n0p_ said:
getprop | grep dexopt
[dalvik.vm.dexopt-flags]: [v=n,o=a,u=y]

Well, same ~950 in Quadrant (CPU/mem only)

998400

Quadrant shows 1048 with o=a and 1065 with o=v. Thats not helping either. At least smartbench shows nearly the same "Games Index". But what do they test with ther productivity index? Calculating PI and mandelbrot is so different?
And there is opera. Its definitly faster with o=a. ~~I'll try to find a web page rendering test.~~

o.k. I get more mixed signals here. Numion tells me that rendering a real world web-page is about 30% faster with o=a (10.5 vs. 14,6 sec). Scragz is 10 % slower with o=a ( 9.7 vs. 10.5 sec).

I'll file it under works for me, but don't try this with miui. Thanks for your help.

Reg. hunderteins

_2bad4u_ · Apr 27, 2012

No problem happy to help

_n0p_ · Apr 27, 2012

hunderteins, this actually lead me to think that you'll try to build highly optimized (armv7, neon, no loop unroll - small cache) libdvm. Am I right?

sinan33 · Apr 28, 2012

_2bad4u_ said:
Tested on miui and it's not working i get into a boot loop.

It works with MIUI too but you have to delete the another line that contains

dalvik.vm.dexopt-flags=m=y

i guess this lines value is overlapping with the other lines "u=y" value.

---------- Post added at 08:11 AM ---------- Previous post was at 08:03 AM ----------

The result of first test is:
Productivity index=716
Gaming index=981

with smart bench 2012

---------- Post added at 08:20 AM ---------- Previous post was at 08:11 AM ----------

The result of second test is:

Productivity index=1074
Gaming index=1569

with smart bench 2011

sinan33 · Apr 28, 2012

DS 5 works faster than ever!

Thank you!

cdzo72 · Apr 28, 2012

hunderteins said:
Hello,

I played with the dalvik.vm.dexopt-flags in /system/build.prop. Basically you control the class-loading and dalvik-vm there.

streakowner proposed in http://xdaforums.com/showpost.php?p=18111740&postcount=644 the setting v=n,o=v. But disabled this with #, in DSC v1.0 the setting is active.
That means: do not verify and optimize mode verified. Found that info in dexopt/OptMain.c

I did some performance measurement with smartbench 2011 and got the value 553 in the first round, and 465 in the second round. That is very bad for an one ghz powered device. Good would > 1000. /data/dalvik-cache takes 74 MiB with this setting.

At the moment I'm running with

dalvik.vm.dexopt-flags=v=n,o=a,u=y

That means: do not verify, optimize all classes and optimize for uniprocessor.

First I was not impressed, because rebooting with a cleaned dalvik-cache took more than double the time span of the o=v setting. logcat is full off lines telling me what can't be optimized and how much time the load took. But after booting I got the impression the system went faster. With smartbench 2011 I get the value 1036 in the first and 1110 in the second round. /data/dalvik-cache takes 74 MiB with this setting.

Can someone confirm my finding?

I don't say you will double the speed of the streak with this setting. But I think you will notice the change. The trade off is the time for the first boot. But I think that is worth is.
So I propose to use the above setting in all gb-based roms for the streak 5.

Have a nice weekend,
hunderteins

Sent from my Dell Streak using Tapatalk 2

Strephon Alkhalikoi · Apr 28, 2012

Testing this now. Opera Mobile is definitely faster, and I honestly think the start time from a wipe of the dalvik cache wasn't any longer than normal.

cdzo72 · Apr 28, 2012

Testing this now. Opera Mobile is definitely faster, and I honestly think the start time from a wipe of the dalvik cache wasn't any longer than normal.

Second that... wiping Dalvik on the Streak seems to be the same for me

Sent from my cm_tenderloin using Tapatalk 2

Edrill · Apr 28, 2012

Same here. Didn't notice any major difference in boot time then usual.

MIUI seems a bit more responsive after this. And I've always found MIUI to be a bit laggy compared to CM and DSC! Gonna observe it for some more time before I can be 100% sure.

sinan33 · Apr 28, 2012

Miui 2.4.20 is not laggy like old versions and it boots very fast.After i edited build.prop it's more responsive and i wrote the results of benches.

_n0p_ · Apr 29, 2012

Just a warning:
When I was experimenting with this settings, i've got caught in situation when CM7 dvm was trying to use /cache for dalvik-cache and failed.
Works fine with Nexus S dvm (as in DSC v7.0)
And I can confirm higher scores in Smartbench.
More, Smartbench somehow gives results, similar to clock value - ~930 on 998400, ~1340 on 1344000, so overclock does matter in this case.

Strephon Alkhalikoi · Apr 29, 2012

So do you recommend using it or going back to the original settings?

_n0p_ · Apr 29, 2012

I sure do.
But just in case keep handy /system/lib/libdvm.so from DSC 7.0 and nandroid backup.
libdvm.so from CM7 has reference to /cache - I wonder if binary change to /data would help.

Dude_Virus · Apr 29, 2012

sinan33 said:
It works with MIUI too but you have to delete the another line that contains

dalvik.vm.dexopt-flags=m=y

i guess this lines value is overlapping with the other lines "u=y" value.

---------- Post added at 08:11 AM ---------- Previous post was at 08:03 AM ----------

The result of first test is:
Productivity index=716
Gaming index=981

with smart bench 2012

---------- Post added at 08:20 AM ---------- Previous post was at 08:11 AM ----------

The result of second test is:

Productivity index=1074
Gaming index=1569

with smart bench 2011

Ok so i need to change only the entry dalvik.vm.dexopt-flags. Below is my buildprop part where the dalvik entries are so just confirm as to what all needs to be changed

#streakOwner's recommended
dalvik.vm.startheapsize=48m
dalvik.vm.heapsize=48m
dalvik.vm.execution-mode=int:jit
dalvik.vm.dexopt-flags=v=n,o=v
dalvik.vm.dexopt-data-only=1
dalvik.vm.lockprof.threshold=150
davlik.vm.verify-bytecode=false
ro.android.kernel.checkjni=0
#wifi.supplicant_scan_interval=120
windowsmgr.max_events_per_sec=80
keyguard.no_require_sim=true

_n0p_ · Apr 30, 2012

Dude_Virus said:
Ok so i need to change only the entry dalvik.vm.dexopt-flags. Below is my buildprop part where the dalvik entries are so just confirm as to what all needs to be changed

#streakOwner's recommended
dalvik.vm.startheapsize=48m
dalvik.vm.heapsize=48m
dalvik.vm.execution-mode=int:jit
dalvik.vm.dexopt-flags=v=n,o=v
dalvik.vm.dexopt-data-only=1
dalvik.vm.lockprof.threshold=150
davlik.vm.verify-bytecode=false
ro.android.kernel.checkjni=0
#wifi.supplicant_scan_interval=120
windowsmgr.max_events_per_sec=80
keyguard.no_require_sim=true

Change
dalvik.vm.dexopt-flags=v=n,o=v

to
dalvik.vm.dexopt-flags=v=n,o=a,u=y
or
dalvik.vm.dexopt-flags=v=n,o=a,u=y,m=y (testing now with register optimization)

Dude_Virus · Apr 30, 2012

_n0p_ said:
Change
dalvik.vm.dexopt-flags=v=n,o=v

to
dalvik.vm.dexopt-flags=v=n,o=a,u=y
or
dalvik.vm.dexopt-flags=v=n,o=a,u=y,m=y (testing now with register optimization)

Thanks for the values....but dont i need to change the cache size to 74 as mentioned in the original post tht the cache goes to 74 mb

_n0p_ · Apr 30, 2012

Dude_Virus said:
Thanks for the values....but dont i need to change the cache size to 74 as mentioned in the original post tht the cache goes to 74 mb

You've took cache size for heap size. That's different things. hunderteins was referring for catalog with built dalvik cache (/data/dalvik-cache)

Proposal to set dalvik-vm optimizing all code

Senior Member

Senior Member

Senior Member

Senior Member

Attachments

Senior Member

Senior Member

Senior Member

Senior Member

Senior Member

Senior Member

Senior Member

Senior Member

Senior Member

Senior Member

Senior Member

Senior Member

Senior Member

Senior Member

Senior Member

Senior Member

Similar threads

Top Liked Posts