On 10/27/2017 10:04 AM, Paul Mackerras wrote:
On Fri, Oct 27, 2017 at 09:38:17AM +0530, Aneesh Kumar K.V wrote:
Hi,
With hash translation mode we always tracked the hash pte slot details in linux
page table.
This occupied space in the linux page table and also limitted our ability to
support
linux features that require additional PTE bits. This series attempt to lift
this
limitation by not tracking slot number in linux page table. We still track slot
details
w.r.t Transparent Hugepage entries because an invalidate there requires us to
go through
all the 256 hash pte slots. So tracking whether hash page table entry is valid
helps us in
avoiding a lot of hcalls there. With THP entries we don't keep slot details in
the primary
linux page table entry but in the second half of page table. Hence tracking
slot details
for THP doesn't take up space in PTE.
Even though we don't track slot, for removing/updating hash page table entry,
PAPR hcalls expect
hash page table slot details. On pseries we find slot using H_READ hcall using
H_READ_4 flags.
This implies an additional 2 hcalls in the updatepp and remove paths. The patch
series also
attempt to limit the impact of this by adding new hcalls that does
remove/update of hash page table
entry using hash value instead of hash page table slot.
Below is the performance numbers observed when running a workload that does the
below sequence
for(5000) {
mmap(128M)
touch every page of 2048 page
munmap()
}
The test is run with address randomization off, swap disabled in both host and
guest.
|------------+----------+---------------+--------------------------+-----------------------|
| iterations | platform | without patch | With series and no hcall | With
series and hcall |
|------------+----------+---------------+--------------------------+-----------------------|
| 1 | powernv | | 50.818343 |
|
| 2 | powernv | | 50.744123 |
|
| 3 | powernv | | 50.721603 |
|
| 4 | powernv | | 50.739922 |
|
| 5 | powernv | | 50.638555 |
|
| 1 | powernv | 51.388249 | |
|
| 2 | powernv | 51.789701 | |
|
| 3 | powernv | 52.240394 | |
|
| 4 | powernv | 51.432255 | |
|
| 5 | powernv | 51.392947 | |
|
|------------+----------+---------------+--------------------------+-----------------------|
| 1 | pseries | | |
123.154394 |
| 2 | pseries | | |
122.253956 |
| 3 | pseries | | |
117.666344 |
| 4 | pseries | | |
117.681479 |
| 5 | pseries | | |
117.735808 |
| 1 | pseries | | 119.424940 |
|
| 2 | pseries | | 117.663078 |
|
| 3 | pseries | | 118.345584 |
|
| 4 | pseries | | 119.620934 |
|
| 5 | pseries | | 119.463185 |
|
| 1 | pseries | 122.810867 | |
|
| 2 | pseries | 115.760801 | |
|
| 3 | pseries | 115.257030 | |
|
| 4 | pseries | 116.617884 | |
|
| 5 | pseries | 117.247036 | |
|
|------------+----------+---------------+--------------------------+-----------------------|
How do we interpret these numbers? Are they times, or speed? Is
larger better or worse?
Sorry for not including the details. They are time in seconds. Test case
is a modified mmap_bench included in powerpc/selftest.
Can you give us the mean and standard deviation for each set of 5
please?
powernv without patch
median= 51.432255
stdev = 0.370835
with patch
median = 50.739922
stdev = 0.06419662
pseries without patch
median = 116.617884
stdev = 3.04531023
with patch no hcall
median = 119.42494
stdev = 0.85874552
with patch and hcall
median = 117.735808
stdev = 2.7624151
-aneesh