On Wed, May 08, 2013 at 11:02:24PM +0200, Artyom Tarasenko wrote:
> On Wed, May 8, 2013 at 12:57 AM, Aurelien Jarno <aurel...@aurel32.net> wrote:
> > On Tue, May 07, 2013 at 11:29:20PM +0200, Artyom Tarasenko wrote:
> >> On Tue, May 7, 2013 at 1:38 PM, Torbjorn Granlund <t...@gmplib.org> wrote:
> >> > The 2nd table of http://gmplib.org/devel/testsystems.html shows all
> >> > emulated systems I am using, most of which are qemu-based.
> >>
> >> Do I read it correct that qemu-system-ppc64 with the slowdown factor
> >> of 33 is ~3 times faster than qemu-system-sparc64 with the slowdown
> >> factor of 96 ?
> >> Do they both use Debian Wheezy guest? You have a remark that ppc64 has
> >> problems with its clock. Was it taken into account when the slowdown
> >> factors were calculated?
> >>
> >
> > Clock or not, it should be noted that qemu-system-sparc64 is undoubtedly
> > slower (at least 5 to 10 times) than qemu-system-{arm,ppc,mips,...} on
> > some type of load like perl scripts.
>
> That's interesting. Actually it should be possible to lauch perl under user
> mode qemu-sparc32plus. Is it possible to launch perl under user mode
> qemu-ppc{32,64} too?
>
> That would allow to understand whether the bad performance has to do
> with TCG or the rest of the system emulation.
I haven't done that yet, but I have run perf top while running perl
script (lintian), on both qemu-system-sparc64 and qemu-system-ppc64. The
results are quite different:
qemu-system-ppc64
-----------------
49,73% perf-10672.map [.] 0x7f7853ab4e0f
13,23% qemu-system-ppc64 [.] cpu_ppc_exec
13,16% libglib-2.0.so.0.3200.4 [.] g_hash_table_lookup
8,18% libglib-2.0.so.0.3200.4 [.] g_str_hash
2,47% qemu-system-ppc64 [.] object_class_dynamic_cast
1,97% qemu-system-ppc64 [.] type_is_ancestor
1,05% libglib-2.0.so.0.3200.4 [.] g_str_equal
0,91% qemu-system-ppc64 [.] ppc_cpu_do_interrupt
0,90% qemu-system-ppc64 [.] object_dynamic_cast_assert
0,79% libc-2.13.so [.] __sigsetjmp
0,62% qemu-system-ppc64 [.] type_get_parent.isra.3
0,58% qemu-system-ppc64 [.] type_get_by_name
0,57% qemu-system-ppc64 [.] qemu_log_mask
0,54% qemu-system-ppc64 [.] object_dynamic_cast
qemu-system-sparc64
-------------------
17,43% perf-8154.map [.] 0x7f6ac10245c8
10,46% qemu-system-sparc64 [.] tcg_optimize
10,36% qemu-system-sparc64 [.] cpu_sparc_exec
6,35% qemu-system-sparc64 [.] tb_flush_jmp_cache
4,75% qemu-system-sparc64 [.] get_physical_address_data
4,45% qemu-system-sparc64 [.] tcg_liveness_analysis
4,35% qemu-system-sparc64 [.] tcg_reg_alloc_op
2,90% qemu-system-sparc64 [.] tlb_flush_page
2,35% qemu-system-sparc64 [.] disas_sparc_insn
2,28% qemu-system-sparc64 [.] get_physical_address_code
2,21% qemu-system-sparc64 [.] tlb_flush
1,64% qemu-system-sparc64 [.] tcg_out_opc
1,22% qemu-system-sparc64 [.] tcg_out_modrm_sib_offset.constprop.41
1,20% qemu-system-sparc64 [.] helper_ld_asi
1,14% qemu-system-sparc64 [.] gen_intermediate_code_pc
1,04% qemu-system-sparc64 [.] helper_st_asi
1,00% qemu-system-sparc64 [.] object_class_dynamic_cast
0,98% qemu-system-sparc64 [.] tb_find_pc
0,94% qemu-system-sparc64 [.] get_page_addr_code
0,92% qemu-system-sparc64 [.] tcg_gen_code_search_pc
0,91% qemu-system-sparc64 [.] tlb_set_page
0,83% qemu-system-sparc64 [.] reset_temp
0,82% qemu-system-sparc64 [.] tcg_reg_alloc_start
The perf-xxxx.map correspond to the code execution. As you can see it's
a lot lower on sparc, while a lot of smaller code generation/mmu code
appears. It's seems that the optimizations have to be focused on the
system part, not the TCG part, at least for now.
A quick look at the MMU seems to show some performance issue here, due
to the split code/data MMU on SPARC64, while the QEMU TLB is a joint
one. As a consequence one can see a lot of ping pong, setting a given
page to read or read/write, then execution, and later read or read/write
again. My guess is that it's related to constants table in the same page
than the code.
It should also be noted that the tcg_optimize starts to take a
non-negligible time, in both cases. The code grew up quite a lot
recently, and it might be time to rework it. It's nice to have optimized
code, but not if the gain is lower than the optimization time.
I am also surprised to see glib code that high on the qemu-system-ppc64
perf report.
--
Aurelien Jarno GPG: 1024D/F1BCDB73
aurel...@aurel32.net http://www.aurel32.net
