Re: 2.6.5-mjb1

Re: 2.6.5-mjb1

[Marcus Hartig]

Hi,

tested with latencytest-0.5.2. Sched stats turned off in the kernel.
Look at "2.6.5 latencytest +aa5 +mm3 as/cfq scheduler" to compare.

---------- 2.6.5-mjb1 AS scheduler ----------
x11.log:
cpu_hz = 2.08792e+09
cpu_load=0.800000  loops per run = 537513
total latency = 1.953125 ms
cpu latency = 1.562500 ms
max diff = 3.03868 ms (overrun 0)
within 1ms = 33088, factor = 99.9788%
within 2ms = 33095, factor = 100%
deviation = 0.0196589

proc.log:
cpu_hz = 2.08792e+09
cpu_load=0.800000  loops per run = 537547
total latency = 1.953125 ms
cpu latency = 1.562500 ms
max diff = 3.05196 ms (overrun 0)
within 1ms = 15385, factor = 99.9805%
within 2ms = 15388, factor = 100%
deviation = 0.0205876

diskwrite.log:
cpu_hz = 2.08792e+09
cpu_load=0.800000  loops per run = 537500
total latency = 1.953125 ms
cpu latency = 1.562500 ms
max diff = 4.40418 ms (overrun 1)
within 1ms = 10989, factor = 99.9273%
within 2ms = 10996, factor = 99.9909%
deviation = 0.0545299

diskcopy.log:
cpu_hz = 2.08792e+09
cpu_load=0.800000  loops per run = 537552
total latency = 1.953125 ms
cpu latency = 1.562500 ms
max diff = 5.08319 ms (overrun 3)
within 1ms = 27949, factor = 99.9321%
within 2ms = 27965, factor = 99.9893%
deviation = 0.0385951

diskread.log:
cpu_hz = 2.08792e+09
cpu_load=0.800000  loops per run = 537564
total latency = 1.953125 ms
cpu latency = 1.562500 ms
max diff = 3.15495 ms (overrun 0)
within 1ms = 28029, factor = 99.9715%
within 2ms = 28037, factor = 100%
deviation = 0.0249104

Best regards,

Marcus

Re: 2.6.5-mjb1

[Martin J. Bligh]

> tested with latencytest-0.5.2. Sched stats turned off in the kernel.
> Look at "2.6.5 latencytest +aa5 +mm3 as/cfq scheduler" to compare.

The main thing that looks different is that loops per run is lower.
What does that mean?
 
> ---------- 2.6.5-mjb1 AS scheduler ----------
> x11.log:
> cpu_hz = 2.08792e+09
> cpu_load=0.800000  loops per run = 537513
> total latency = 1.953125 ms
> cpu latency = 1.562500 ms
> max diff = 3.03868 ms (overrun 0)
> within 1ms = 33088, factor = 99.9788%
> within 2ms = 33095, factor = 100%
> deviation = 0.0196589
> 
> proc.log:
> cpu_hz = 2.08792e+09
> cpu_load=0.800000  loops per run = 537547
> total latency = 1.953125 ms
> cpu latency = 1.562500 ms
> max diff = 3.05196 ms (overrun 0)
> within 1ms = 15385, factor = 99.9805%
> within 2ms = 15388, factor = 100%
> deviation = 0.0205876
> 
> diskwrite.log:
> cpu_hz = 2.08792e+09
> cpu_load=0.800000  loops per run = 537500
> total latency = 1.953125 ms
> cpu latency = 1.562500 ms
> max diff = 4.40418 ms (overrun 1)
> within 1ms = 10989, factor = 99.9273%
> within 2ms = 10996, factor = 99.9909%
> deviation = 0.0545299
> 
> diskcopy.log:
> cpu_hz = 2.08792e+09
> cpu_load=0.800000  loops per run = 537552
> total latency = 1.953125 ms
> cpu latency = 1.562500 ms
> max diff = 5.08319 ms (overrun 3)
> within 1ms = 27949, factor = 99.9321%
> within 2ms = 27965, factor = 99.9893%
> deviation = 0.0385951
> 
> diskread.log:
> cpu_hz = 2.08792e+09
> cpu_load=0.800000  loops per run = 537564
> total latency = 1.953125 ms
> cpu latency = 1.562500 ms
> max diff = 3.15495 ms (overrun 0)
> within 1ms = 28029, factor = 99.9715%
> within 2ms = 28037, factor = 100%
> deviation = 0.0249104
> 
> Best regards,
> 
> Marcus
> 
> 
> 

Re: 2.6.5-mjb1

[Marcus Hartig]

Martin J. Bligh wrote:

> The main thing that looks different is that loops per run is lower.
> What does that mean?

Hmm. Sorry, here seems to be an error with my first run of the -mjb1 
bench. Now, in the second try it has the loops running as the others, I 
think. And a good performance...

x11.log:
cpu_hz = 2.0885e+09
cpu_load=0.800000  loops per run = 645177
total latency = 1.953125 ms
cpu latency = 1.562500 ms
max diff = 3.06248 ms (overrun 0)
within 1ms = 38919, factor = 99.9794%
within 2ms = 38927, factor = 100%
deviation = 0.0216797

proc.log:
cpu_hz = 2.0885e+09
cpu_load=0.800000  loops per run = 645114
total latency = 1.953125 ms
cpu latency = 1.562500 ms
max diff = 2.96921 ms (overrun 0)
within 1ms = 15422, factor = 99.987%
within 2ms = 15424, factor = 100%
deviation = 0.0223192

diskwrite.log:
cpu_hz = 2.0885e+09
cpu_load=0.800000  loops per run = 645226
total latency = 1.953125 ms
cpu latency = 1.562500 ms
max diff = 4.64053 ms (overrun 2)
within 1ms = 13345, factor = 99.9176%
within 2ms = 13354, factor = 99.985%
deviation = 0.0517744

diskcopy.log:
cpu_hz = 2.0885e+09
cpu_load=0.800000  loops per run = 644902
total latency = 1.953125 ms
cpu latency = 1.562500 ms
max diff = 3.45113 ms (overrun 0)
within 1ms = 40427, factor = 99.9357%
within 2ms = 40453, factor = 100%
deviation = 0.0407751

diskread.log:
cpu_hz = 2.0885e+09
cpu_load=0.800000  loops per run = 644972
total latency = 1.953125 ms
cpu latency = 1.562500 ms
max diff = 3.2573 ms (overrun 0)
within 1ms = 38383, factor = 99.9401%
within 2ms = 38406, factor = 100%
deviation = 0.0324159


Marcus

2.6.5-mjb1

[Martin J. Bligh]

The patchset is meant to be pretty stable, not so much a testing ground.
Main differences from mainline are:

1. Better performance & resource consumption, particularly on larger machines.
2. Diagnosis tools (kgdb, early_printk, etc).
3. Kexec support.
4. ivtv drivers

I'd be very interested in feedback from anyone willing to test on any 
platform, however large or small.

ftp://ftp.kernel.org/pub/linux/kernel/people/mbligh/2.6.5/patch-2.6.5-mjb1.bz2

Since 2.6.5-rc3-mjb2 (~ = changed, + = added, - = dropped)

Notes: 

-----------------------------------------------------------------------

Now in Linus' tree:


Dropped:

New:

+ aio-retry
+ 4g4g-aio-hang-fix
+ aio-retry-elevated-refcount
+ aio-splice-runlist
+ aio-wait-page
+ aio-fs_read
+ aio-upfront-readahead
+ O_SYNC-speedup
+ aio-O_SYNC
+ gang_lookup_next
+ aio-gang_lookup-fix
+ aio-O_SYNC-short-write-fix.patch
+ aio-cancel-fix
+ aio-read-immediate
+ aio-pipe
+ aio-context-switch
+ aio-putioctx-flushworkqueue
+ aio-poll
	AIO filesystem support				Suparna et al.
+ per_task_TUB_PPC32					Anton 
	Enable per-task TASK_UMAPPED_BASE on PPC32
+ jbd_journal_speedup					Andrew Morton
	ext3 fs speedup
+ amd64_sched						Darren Hart
	Make AMD scheduler behave as flat SMP, not NUMA

Pending:
per_node_rss
local_balance_exec
reluctance in cross-node balance (less_bouncy)
sched tunables patch
emulex update
NUMA membinding API
x86_64 update
config_numasched
sched tunables (reinstante)
list_of_lists
Child runs first (akpm)
Netdump

Present in this patch:

-mjb						Martin J. Bligh
	Add a tag to the makefile

kgdb						Various
	Stolen from akpm's 2.6.0-mm1, includes fixes

kgdboe_netpoll					Matt Mackall et al.
	Kgdb over ethernet support that works with the netpoll infrastructure

kgdboe_build_fix				Andrew Morton
	Fix kgdboe stuff so non-ia32 platforms build

kgdb_x86_64					Jim Houston
	Support kgdb on x86_64

kgdb_gdb6_patches				Jim Houston
	Patches for gdb to support kgdb on x86_64, under scripts/kgdb/

ppc64_reloc_hide				Anton Blanchard / Paul Mackerras
	PPC 64 fixups

spinlock_inlining				Andrew Morton & Martin J. Bligh
	Inline spinlocks for profiling. Made into a ugly config option by me.

lockmeter					John Hawkes / Hanna Linder
	Locking stats.

lockmeter_ia64					Ray Bryant
	Add a config option for lockmeter on ia64

oops_dump_preceding_code			Andrew Morton
	dump opcodes preceding and after the offending EIP.

scheduler_2.6.5_rc3_mm1				Nick Piggin
	sched_domains code

4k_stacks					Arjan / Dave H / bcrl
	Provide an option to use 4k kernel stacks instead of 8k

confighz					Andrew Morton / Dave Hansen
	Make HZ a config option of 100 Hz or 1000 Hz

config_page_offset				Dave Hansen / Andrea
	Make PAGE_OFFSET a config option

numameminfo					Martin Bligh / Keith Mannthey
	Expose NUMA meminfo information under /proc/meminfo.numa

partial_objrmap					Dave McCracken
	Object based rmap for filebacked pages.

anon_mm2					Hugh Dickins
anon_mm3					Hugh Dickins
anon_mm4					Hugh Dickins
anon_mm5					Hugh Dickins
anon_mm6					Hugh Dickins
anon_mm7					Hugh Dickins
anon_mm8					Hugh Dickins
	Object based rmap for anonymous memory

4g4g						Ingo Molnar
	Provide a 4G/4G user/kernel split for 32 bit memory lushes.

4g_zap_low_mappings					Martin Lorenz
	stop zap_low_mappings from being __init

4g4g_locked_copy					Dave McCracken
	Fix locking bug in 4/4 split

disable preempt					Martin J. Bligh
	I broke preempt somehow, temporarily disable it to stop accidents

aiofix2						Mingming Cao
	fixed a bug in ioctx_alloc()

percpu_real_loadavg				Dave Hansen / Martin J. Bligh
	Tell me what the real load average is, and tell me per cpu.

gfp_node_strict					Dave Hansen
	Add a node strict binding as a gfp mask option

irqbal_fast					Adam Litke
	Balance IRQs more readily

kcg						Adam Litke
	Acylic call graphs from the kernel. Wheeeeeeeeeeeee!

kcg_gcc_detect					Adam Litke
	Detect older gcc versions that don't work with mcount, and crap out

numa_mem_equals 				Dave Hansen
	mem= command line parameter NUMA awareness.

autoswap					Con Kolivas
	Auto-tune swapiness

emulex driver					Emulex
	Driver for emulex fiberchannel cards

multiple_emulex					Mike Anderson
	Allow multiple Emulex cards

protocol254					Paul Mackerras / Omkhar 
	Allow protocol 254

slabtune					Dave McCracken
	Take slab in bigger bites on larger machines

topdown						Bill Irwin
	Turn userspace upside down for fun & profit

stacktrace					Adam Litke
	Stack backtracing via frame pointers

fasync_lock_rcu					Manfred Spraul
	Use RCU for fasync_lock

aio-retry
4g4g-aio-hang-fix
aio-retry-elevated-refcount
aio-splice-runlist
aio-wait-page
aio-fs_read
aio-upfront-readahead
O_SYNC-speedup
aio-O_SYNC
gang_lookup_next
aio-gang_lookup-fix
aio-O_SYNC-short-write-fix.patch
aio-cancel-fix
aio-read-immediate
aio-pipe
aio-context-switch
aio-putioctx-flushworkqueue
aio-poll
	AIO filesystem support				Suparna et al.

kexec						Eric Biederman et al.
	Exec a kernel for breakfast today.

lockmeter_notsc					Martin J. Bligh
	Lockmeter does not require CONFIG_X86_TSC.

tiocgdev						Gerd Knorr

vma_statistics						Martin J. Bligh
	Provide per VMA stats

per_task_TUB						Adam Litke
	Per task TASK_UNMAPPED_BASE

per_task_TUB_PPC32					Anton 
	Enable per-task TASK_UMAPPED_BASE on PPC32

implicit_hugetlb					Adam Litke
	Implicit allocation of huge pages

hugetlb_dyn_as						Adam Litke
	Dynamic huge pages.

irq_vector						James Cleverdon
	Fix irq vector limits for Summit

ivtv						Kevin Thayer / Steven Fuerst
	Driver for ivtv (includes Hauppauge PVR 250 / 350)
	Written by Kevin Thayer, ported to 2.6 by Steven Fuerst
	Version 0.1.9

sysfs_backing_store1					Maneesh Soni
sysfs_backing_store2					Maneesh Soni
sysfs_backing_store3					Maneesh Soni
sysfs_backing_store4					Maneesh Soni
sysfs_backing_store5					Maneesh Soni
sysfs_backing_store6					Maneesh Soni
	Make sysfs more efficient in its usage of lowmem

vgtod1							John Stultz
vgtod2							John Stultz
vgtod3							John Stultz
	Vsyscall gettimeofday for ia32

physnode_map						Martin J. Bligh
	Hack around problem of missing area in physnode_map

sched_tunables						R. Love / Darren Hart
	Provide sched tunables to play with on a rainy day.

zone_gap						Andy Whitcroft
	Fix up the gap between ZONE_NORMAL and ZONE_HIGHMEM on NUMA.

max_mp_busses						James Cleverdon
	Increase MAX_MP_BUSSES

schedstats						Rick Lindsley
	Provide lotsa scheduler statistics

schedstats-tools					Rick Lindsley
	Grub around in lotsa scheduler statistics

jbd_journal_speedup					Andrew Morton
	ext3 fs speedup

amd64_sched						Darren Hart
	Make AMD scheduler behave as flat SMP, not NUMA