This is the source-code for Contiguity Aware Paging. An add-on on Linux memory manager to allocate target pages and create larger-than-a-page contiguous mappings. These mappings can be used by novel Address Translation Hardware to speedup translation. CA paging is presented on Enhancing and Exploiting Contiguity for Fast Memory Virtualization (ISCA '20).
Currently CA paging is optimized to work on top of Transparent Huge Pages.
Therefore, for better performance configure Linux with THP always on.
CONFIG_TRANSPARENT_HUGEPAGE=y
CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS=y
To use the VM Introspection tool (VMI) and collect statistics (see below)
Kernel Address Space Layout Randomization (KASLR) must be turned off
both for the guest and the host OS.
CONFIG_RANDOMIZE_BASE=n
There is a working .config file on this repo.
CA paging is turned on for a process (process's page faults and memory mapped files will trigger contiguous allocations) with a system call (NR 337).
There is a utility C-program provided to utilize CA paging easily.
Go to the directory tools/CAPaging_scripts/run/native/:
Compile the code:
gcc -o capaging capaging.c
Run a process with CA paging on:
./capaging command <command to run>
(e.g. ./capaging command ./XSBench -t 1 -s XL -l 64 -G unionized)
You can also run a process with CA paging by including in its source code
the following line of code:
syscall(337, NULL, 0, 0)
There is a utility C-program provided to collect statistics for
the contiguity of a process's mappings (regardless using CA paging)
based on page-collect.c.
Go to the directory tools/CAPaging_scripts/collect-statistic/native/:
Compile the code:
make
Collect statistics for a process:
./page-collect -p <pid of the process> -o <output file name>
Output:
The output file contains contiguity statistics for the following types of mappings:
- VMA TLB: each entry is a process's VMA area.
- Virtual TLB: each entry is a set of contiguous virtual pages that are backed by physical memory.
- Anchor TLB: each entry is a set of contiguous virtual pages mapped to contiguous physical pages, where the starting virtual address is aligned to an anchor size as described in Hybrid TLB Coalescing.
- Range TLB: each entry is a set of contiguous virtual pages mapped to contiguous physical pages with no alignment restrictions as described in Redundant Memory Mappings. As there are no restrictions, this captures all the contiguous virtual-to-physical mappings of the process intact.
- SpOT: each entry is a set of page mappings whith the same Offset (virtual-physical address), without the requirement that the virtual addresses are contiguous as described in Enhancing and Exploiting Contiguity.
An output example
----------
VMA_TLB
----------
total_VMA_TLB_entries: 35
total_VMA_TLB_coverage: 120018 (MB)
32 entries coverage: 100.01% (30724672 4K pages)
64 entries coverage: 0.00% (0 4K pages)
128 entries coverage: 0.00% (0 4K pages)
256 entries coverage: 0.00% (0 4K pages)
Number of entries for at least 80% coverage: 1 (exact coverage 100.00%)
Number of entries for at least 90% coverage: 1 (exact coverage 100.00%)
Number of entries for at least 99% coverage: 1 (exact coverage 100.00%)
Virtual_TLB
total_Virtual_TLB_entries: 41
total_Virtual_TLB_coverage: 120006 (MB)
32 entries coverage: 100.00% (30721560 4K pages)
64 entries coverage: 0.00% (0 4K pages)
128 entries coverage: 0.00% (0 4K pages)
256 entries coverage: 0.00% (0 4K pages)
Number of entries for at least 80% coverage: 1 (exact coverage 100.00%)
Number of entries for at least 90% coverage: 1 (exact coverage 100.00%)
Number of entries for at least 99% coverage: 1 (exact coverage 100.00%)
----------
Anchor_TLB
----------
total_Anchor_TLB_entries: 3875
total_Anchor_TLB_coverage: 120006 (MB)
32 entries coverage: 13.65% (4194304 4K pages)
64 entries coverage: 27.31% (8388608 4K pages)
128 entries coverage: 53.88% (16551936 4K pages)
256 entries coverage: 81.18% (24940544 4K pages)
Number of entries for at least 80% coverage: 251 (exact coverage 80.12%)
Number of entries for at least 90% coverage: 298 (exact coverage 90.14%)
Number of entries for at least 99% coverage: 1060 (exact coverage 99.00%)
----------
Range_TLB
----------
total_Range_TLB_entries: 532
total_Range_TLB_coverage: 120006 (MB)
32 entries coverage: 100.00% (30720556 4K pages)
64 entries coverage: 100.00% (30721070 4K pages)
128 entries coverage: 100.00% (30721165 4K pages)
256 entries coverage: 100.00% (30721293 4K pages)
Number of entries for at least 80% coverage: 8 (exact coverage 80.37%)
Number of entries for at least 90% coverage: 11 (exact coverage 90.08%)
Number of entries for at least 99% coverage: 17 (exact coverage 99.39%)
----------
SpOT
----------
total_SpOT_entries: 518
total_SpOT_coverage: 120006 (MB)
32 entries coverage: 100.00% (30720707 4K pages)
64 entries coverage: 100.00% (30721095 4K pages)
128 entries coverage: 100.00% (30721179 4K pages)
256 entries coverage: 100.00% (30721307 4K pages)
Number of entries for at least 80% coverage: 8 (exact coverage 80.43%)
Number of entries for at least 90% coverage: 11 (exact coverage 90.13%)
Number of entries for at least 99% coverage: 16 (exact coverage 99.44%)
----------
working_set
----------
total_present_pages: 30721569
4K pages: 2081
2M pages: 59999
total_present_working_set: 120006 (MB)
To create guest virtual to host physical (gVA-to-hPA) contiguous mappings CA paging must be turned on for:
- guest OS: the guest application (creates contiguous gVA-to-gPA mappings)
- host OS: the hypervisor (creates contiguous gPA-to-hPA mappings)
Example:
- In the host launch qemu with CA paging:
./capaging command qemu-system-x86_64 -enable-kvm -m 256G disk.img
- In the guest run the application with CA paging:
./capaging command ./XSBench -t 1 -s XL -l 64 -G unionized
To enable statistics collection for a guest process's gVA-to-hPA mappings from the host OS we have developped a VM Introspection tool. It exposes in the host OS a proc interface with functionality similar to /proc/maps and /proc/pagemap but for a target guest process.
Specifically in the host OS:
- /proc/VMI/<qemu pid>/<guest process pid>/comm: prints the guest process command name
- /proc/VMI/<qemu pid>/<guest process pid>/maps: prints the guest process contiguous virtual memory regions (gVA)
- /proc/VMI/<qemu pid>/<guest process pid>/pagemap: is an interface that enables getting the hPA translation for a process's gVA.
There are utility C-programs provided to utilize the VMI. Both the guest and the host must run with the same kernel and the KASLR off.
CONFIG_RANDOMIZE_BASE=n
Go to the directory tools/CAPaging_scripts/run/virtual/:
To register a guest process for VMI monitoring run inside the guest:
CA paging off (default demand paging):
gcc -o vmi vmi.c
./vmi command <command to run>
(e.g. ./vmi command ./XSBench -t 1 -s XL -l 64 -G unionized)
CA paging on:
gcc -o vmi-capaging vmi-capaging.c
./vmi-capaging command <command to run>
(e.g. ./vmi-capaging command ./XSBench -t 1 -s XL -l 64 -G unionized)
Go to the directory tools/CAPaging_scripts/collect-statistics/VMI/:
Run in the host:
./page-collect -p <pid of the guest process> -q <qemu pid> -o <output file name>
The output is the same as explained in the native execution section.
- /proc/capaging_contiguity_map: the current contiguity map of the system
- /proc/capaging/
- failure: number and reason of failed target page allocations
- success: number and category of succesfull target page allocations
- total_alloc: total number of capaging target allocation attempts