Analysis of the VM Performance Deterioration When Running memcpy to Copy 1,000 Bytes in the x86_64 Environment
1. Problem Background
1.1 Symptom
When memcpy 1k is executed in the x86_64 environment, the virtual machine (VM) performance is 40 times lower than that of the physical machine (PM).
1.2 Software Information
| Item | Version |
|---|---|
| OS | openEuler 20.03 LTS |
| Kernel | 4.19.90-2003.4.0.0036.oe1.x86_64 |
| Glibc | 2.28 |
| GCC | 7.3.0 |
2. Conclusion and Solution
2.1 Conclusion
The hyper-threading function is not enabled in the XML file for starting the VM. As a result, the memcpy L3 cache threshold on the PM is different from that on the VM, causing the performance difference.
2.2 Solution
Method 1: Enabling Hyper-Threading for the VM
<cpu mode='host-passthrough' check='none'>
...
<topology sockets='2' cores='4' threads='2'/>
<feature policy='require' name='topoext'/>
</cpu>Method 2: Adjusting the memcpy Threshold
The following configuration is recommended by the glibc community:
# export GLIBC_TUNABLES=glibc.tune.x86_non_temporal_threshold=$(($(getconf LEVEL3_CACHE_SIZE) * 3 / 4))3 Overview of the memcpy Algorithm
In glibc-2.28, memcpy and memove share the same set of logic. The implementation algorithm is briefly described in the glibc source code.
sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
/* memmove/memcpy/mempcpy is implemented as:
1. Use overlapping load and store to avoid branch.
2. Load all sources into registers and store them together to avoid
possible address overlap between source and destination.
3. If size is 8 * VEC_SIZE or less, load all sources into registers
and store them together.
4. If address of destination > address of source, backward copy
4 * VEC_SIZE at a time with unaligned load and aligned store.
Load the first 4 * VEC and last VEC before the loop and store
them after the loop to support overlapping addresses.
5. Otherwise, forward copy 4 * VEC_SIZE at a time with unaligned
load and aligned store. Load the last 4 * VEC and first VEC
before the loop and store them after the loop to support
overlapping addresses.
6. If size >= __x86_shared_non_temporal_threshold and there is no
overlap between destination and source, use non-temporal store
instead of aligned store. */As described in item 6, if __x86_shared_non_temporal_threshold is exceeded, the non-temporal store is used to replace the aligned store, which causes the performance deterioration.
4 Execution Logic
Before the process is started in the x86 environment, a series of threshold initialization operations are performed. The involved threshold initialization operations are as follows:
sysdeps/x86/cacheinfo.c
533 /* A value of 0 for the HTT bit indicates there is only a single
534 logical processor. */
535 if (HAS_CPU_FEATURE (HTT))
536 {
...
Compute threads
...
693 }
...
781 /* The large memcpy micro benchmark in glibc shows that 6 times of
782 shared cache size is the approximate value above which non-temporal
783 store becomes faster on a 8-core processor. This is the 3/4 of the
784 total shared cache size. */
785 __x86_shared_non_temporal_threshold
786 = (cpu_features->non_temporal_threshold != 0
787 ? cpu_features->non_temporal_threshold
788 : __x86_shared_cache_size * threads * 3 / 4);It shows that the value of __x86_shared_non_temporal_threshold of the VM is 0 when hyper-threading is not enabled, and that of the PM is __x86_shared_cache_size * threads * 3/4. When the memcpy 1k operation is performed, the branch to be executed is determined based on the following logic. The logic of the VM is different from that of the PM after this operation.
sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
455 #if (defined USE_MULTIARCH || VEC_SIZE == 16) && IS_IN (libc)
456 /* Check non-temporal store threshold. */
457 cmpq __x86_shared_non_temporal_threshold(%rip), %rdx
458 ja L(large_backward)
459 #endifThe logic is as follows:
PM logic
sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
460 L(loop_4x_vec_backward):
461 /* Copy 4 * VEC a time backward. */
462 VMOVU (%rcx), %VEC(0)
463 VMOVU -VEC_SIZE(%rcx), %VEC(1)
464 VMOVU -(VEC_SIZE * 2)(%rcx), %VEC(2)
465 VMOVU -(VEC_SIZE * 3)(%rcx), %VEC(3)
466 subq $(VEC_SIZE * 4), %rcx
467 subq $(VEC_SIZE * 4), %rdx
468 VMOVA %VEC(0), (%r9)
469 VMOVA %VEC(1), -VEC_SIZE(%r9)
470 VMOVA %VEC(2), -(VEC_SIZE * 2)(%r9)
471 VMOVA %VEC(3), -(VEC_SIZE * 3)(%r9)
472 subq $(VEC_SIZE * 4), %r9
473 cmpq $(VEC_SIZE * 4), %rdx
474 ja L(loop_4x_vec_backward)
475 /* Store the first 4 * VEC. */
476 VMOVU %VEC(4), (%rdi)
477 VMOVU %VEC(5), VEC_SIZE(%rdi)
478 VMOVU %VEC(6), (VEC_SIZE * 2)(%rdi)
479 VMOVU %VEC(7), (VEC_SIZE * 3)(%rdi)
480 /* Store the last VEC. */
481 VMOVU %VEC(8), (%r11)
482 VZEROUPPER
483 retVM logic
sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
528 L(loop_large_backward):
529 /* Copy 4 * VEC a time backward with non-temporal stores. */
530 PREFETCH_ONE_SET (-1, (%rcx), -PREFETCHED_LOAD_SIZE * 2)
531 PREFETCH_ONE_SET (-1, (%rcx), -PREFETCHED_LOAD_SIZE * 3)
532 VMOVU (%rcx), %VEC(0)
533 VMOVU -VEC_SIZE(%rcx), %VEC(1)
534 VMOVU -(VEC_SIZE * 2)(%rcx), %VEC(2)
535 VMOVU -(VEC_SIZE * 3)(%rcx), %VEC(3)
536 subq $PREFETCHED_LOAD_SIZE, %rcx
537 subq $PREFETCHED_LOAD_SIZE, %rdx
538 VMOVNT %VEC(0), (%r9)
539 VMOVNT %VEC(1), -VEC_SIZE(%r9)
540 VMOVNT %VEC(2), -(VEC_SIZE * 2)(%r9)
541 VMOVNT %VEC(3), -(VEC_SIZE * 3)(%r9)
542 subq $PREFETCHED_LOAD_SIZE, %r9
543 cmpq $PREFETCHED_LOAD_SIZE, %rdx
544 ja L(loop_large_backward)
545 sfence
546 /* Store the first 4 * VEC. */
547 VMOVU %VEC(4), (%rdi)
548 VMOVU %VEC(5), VEC_SIZE(%rdi)
549 VMOVU %VEC(6), (VEC_SIZE * 2)(%rdi)
550 VMOVU %VEC(7), (VEC_SIZE * 3)(%rdi)
551 /* Store the last VEC. */
552 VMOVU %VEC(8), (%r11)
553 VZEROUPPER
554 ret5 Instruction Difference Analysis
According to the preceding description, the biggest difference between the execution logic of the PM and that of the VM lies in the mov instruction. The definition of the mov instruction is as follows:
sysdeps/x86_64/memmove.S
23 #define PREFETCHNT prefetchnta
24 #define VMOVNT movntdq
25 /* Use movups and movaps for smaller code sizes. */
26 #define VMOVU movups
27 #define VMOVA movapsThe PM logic uses the movaps instruction, which features 16-byte alignment. The VM logic uses the movntdq instruction, which bypasses the main cache. See the following description.
https://stackoverflow.com/questions/14106477/how-do-non-temporal-instructions-work
The streaming read/write with non-temporal hints are typically used to reduce cache pollution (often with WC memory). The idea is that a small set of cache lines are reserved on the CPU for these instructions to use. Instead of loading a cache line into the main caches, it is loaded into this smaller cache.The comment supposes the following behavior (but I cannot find any references that the hardware actually does this, one would need to measure or a solid source and it could vary from hardware to hardware): - Once the CPU sees that the store buffer is full and that it is aligned to a cache line, it will flush it directly to memory since the non-temporal write bypasses the main cache.
As described, the movntdq instruction stores data in the memory by bypassing the main cache. Therefore, the performance of the movntdq instruction is inferior to that of the movaps instruction.
After communicating with the community, we know that the community adopts a compromise strategy. For the memcpy operation of large data blocks, if the L3 cache is used, the memcpy performance can be improved, but the performance of the entire system is affected. Therefore, the threshold is specified.
https://sourceware.org/pipermail/libc-alpha/2021-January/121510.html
> The performance of memcpy 1024 has recovered. However, there is performance
> reduce in host. This is test result (cycle):
>
> memcpy_10 memcpy_1k memcpy_10k memcpy_1m memcpy_10m
> before backport 8 34 187 130848 2325409
> after backport 8 34 182 515156 5282603
> Performance improvement 0.00% 0.00% 2.67% -293.71% -127.17%I think this is expected because the large copies no longer stay within the cache. This is required to avoid blowing away the entire cache contents for such large copies, negatively impacting whole system performance. This will of course not show up in a micro-benchmark.
6 After Threshold Change
According to the preceding analysis, the default threshold of the VM is 0. A verification test is performed on the VM and PM based on the recommended configuration of the community. The result (unit: number of cycles) is as follows:
| Physical Machine | memcpy_10 | memcpy_1k | memcpy_10k | memcpy_1M | memcpy_10M |
|---|---|---|---|---|---|
| Before the configuration | 8 | 34 | 187 | 130848 | 2325409 |
| After the configuration | 8 | 34 | 182 | 515156 | 5282603 |
| Performance | 0.00% | 0.00% | 2.67% | -293.71% | -127.17% |
| Virtual Machine | memcpy_10 | memcpy_1k | memcpy_10k | memcpy_1M | memcpy_10M |
|---|---|---|---|---|---|
| Before the configuration | 8 | 1269 | 4555 | 523740 | 5304273 |
| After the configuration | 8 | 35 | 183 | 509297 | 5260913 |
| Performance | 0.00% | 97.24% | 95.98% | 2.76% | 0.82% |
| Compare the statistics before and after the configuration of the VM and PM. It is found that the performance of the VM is the same as that of the PM after the threshold is changed. In addition, for the PM, the threshold is changed from __x86_shared_cache_size * threads * 3/4 to __x86_shared_cache_size * 3/4. The threshold decreases, and the movntdq instruction is more likely to be used. Therefore, the PM performance decreases when the data block is greater than or equal to 1 MB. |
