Cacheline demote to improve performance #11101

wxue1 · 2023-04-19T09:27:40Z

Once code is emitted to JIT buffer, hint the hardware to demote the corresponding cache lines to more distant level so other CPUs can access them more quickly.
This gets nearly 1% performance gain on our workload.

dstogov · 2023-04-19T14:09:07Z

I think, first we should check if CLDEMOTE is available, using CPUID.
At second, I'm not sure if we should always use it. Function JIT may produce more than 10M of code at single request.
Only the iteration over this 10M may take time, then this 10M won't fit into cache, evict something useful, and finally it's not necessary need by the other cores.

Can you demonstrate CLDEMOTE usage in some other JIT engines?

@wxue1 we have been working on a new generation JIT engine for a while. See https://github.com/dstogov/ir
You are welcome there :)

wxue1 · 2023-04-20T01:53:09Z

CLDEMOTE

Thanks and I will check it based on your comment.

@wxue1 we have been working on a new generation JIT engine for a while. See https://github.com/dstogov/ir You are welcome there :)

Great, more IR optimizations could be added then. I'll deep dive into this later.

Another thing is that could you help to review PR #10897 since I updated it.

dstogov · 2023-04-20T07:24:12Z

Another thing is that could you help to review PR #10897 since I updated it.

Yeah. I saw. Sorry, I'm limited in time now and that PR is not simple. I'll take a careful look on Monday or Tuesday.

wxue1 · 2023-04-24T01:27:31Z

I think, first we should check if CLDEMOTE is available, using CPUID. At second, I'm not sure if we should always use it. Function JIT may produce more than 10M of code at single request. Only the iteration over this 10M may take time, then this 10M won't fit into cache, evict something useful, and finally it's not necessary need by the other cores.

On processors which do not support the CLDEMOTE instruction (including legacy hardware) the instruction will be treated as a NOP. (from the software development manual)

I update the code to support tracing pattern only.

CLDEMOTE haven't taken by other JIT engine, but DPDK, LLVM and other repos support that.

dstogov · 2023-04-24T11:22:21Z

I think, first we should check if CLDEMOTE is available, using CPUID. At second, I'm not sure if we should always use it. Function JIT may produce more than 10M of code at single request. Only the iteration over this 10M may take time, then this 10M won't fit into cache, evict something useful, and finally it's not necessary need by the other cores.

On processors which do not support the CLDEMOTE instruction (including legacy hardware) the instruction will be treated as a NOP. (from the software development manual)

10M/64 of NOPs :(
See Zend/zend_cpuinfo.h. It shouldn't be a big problem to detect CLDEMOTE support.

CLDEMOTE haven't taken by other JIT engine, but DPDK, LLVM and other repos support that.

I see they provide ability to generate CLDEMOTE but don't use CLDEMOTE their selves.

Few years ago Intel engineers suggested replacing non temporal stores into opcache memory by regular ones.
See 68185bafbe2
Your current suggestion is a bit opposite or something in the middle.

dstogov · 2023-04-24T12:57:20Z

Oh, I think I finally understood how CLDEMOTE may improve the overall speed. It hints CPU to evict data from L1 and L2 caches into L3. And than the data in L3 may be reused by other CPU cores.

wxue1 · 2023-04-25T03:50:14Z

Oh, I think I finally understood how CLDEMOTE may improve the overall speed. It hints CPU to evict data from L1 and L2 caches into L3. And than the data in L3 may be reused by other CPU cores.

Yes, let me check Zend/zend_cpuinfo.h to support CLDEMOTE.

dstogov · 2023-04-25T06:31:47Z

You may also try to use CLDEMOTE in ext/opcache/ZendAccelerator.c at the end of cache_script_in_shared_memory() to demote the cached script. Of course, this requites some patch adjustment.

	shared_cacheline_demote(new_persistent_script->mem, new_persistent_script->size);

wxue1 · 2023-04-25T07:55:42Z

You may also try to use CLDEMOTE in ext/opcache/ZendAccelerator.c at the end of cache_script_in_shared_memory()

Yeah, I tried before. But PHP store script structure in the hash table which is scattered. CLDEMOTE work effect is not obvious

Once code is emitted to JIT buffer, hint the hardware to demote the corresponding cache lines to more distant level so other CPUs can access them more quickly. This gets nearly 1% performance gain on our workload. Signed-off-by: Xue,Wang <[email protected]> Signed-off-by: Tao,Su <[email protected]> Signed-off-by: Hu,chen <[email protected]>

wxue1 · 2023-05-12T09:25:09Z

Oh, I think I finally understood how CLDEMOTE may improve the overall speed. It hints CPU to evict data from L1 and L2 caches into L3. And than the data in L3 may be reused by other CPU cores.

Yes, let me check Zend/zend_cpuinfo.h to support CLDEMOTE.

I updated the patch and checked CPUINFO whether cldemote instruction is available

dstogov · 2023-05-15T06:53:30Z

ext/opcache/jit/zend_jit.c

+	/* hint to the hardware to push out the cache line that contains the linear address */
+#if ZEND_JIT_SUPPORT_CLDEMOTE
+	if (cpu_support_cldemote && JIT_G(trigger) == ZEND_JIT_ON_HOT_TRACE) {
+		shared_cacheline_demote((uintptr_t)entry, size);
+	}
+#endif


What is the reason to do CLDEMOTE only for TRACEs?

What is the reason to do CLDEMOTE only for TRACEs?

You mentioned that Function JIT may produce more than 10M of code at a single request and CLDEMOTE takes time. So we support tracing JIT.

wxue1 · 2023-05-15T08:05:25Z

Thanks! I will focus on my other inline function patch and PHP9 IR next.

wxue1 requested review from dstogov and iluuu1994 as code owners April 19, 2023 09:27

github-actions bot added the Extension: opcache label Apr 19, 2023

wxue1 force-pushed the cacheline_demote branch 2 times, most recently from 1435f91 to 00c6d3d Compare April 23, 2023 09:03

wxue1 force-pushed the cacheline_demote branch from 00c6d3d to d357c59 Compare May 12, 2023 07:39

github-actions bot added the Category: Engine label May 12, 2023

dstogov reviewed May 15, 2023

View reviewed changes

dstogov approved these changes May 15, 2023

View reviewed changes

dstogov merged commit 6bd5464 into php:master May 15, 2023

pronskiy mentioned this pull request Jun 6, 2023

Add Roundup #13 ThePHPF/thephp.foundation#90

Merged

stkeke deleted the cacheline_demote branch August 10, 2023 08:22

PeterYang12 mentioned this pull request Nov 29, 2023

Cacheline demote optimization for JITTed code is missing #12825

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Cacheline demote to improve performance #11101

Cacheline demote to improve performance #11101

wxue1 commented Apr 19, 2023

dstogov commented Apr 19, 2023

wxue1 commented Apr 20, 2023 •

edited

Loading

dstogov commented Apr 20, 2023

wxue1 commented Apr 24, 2023 •

edited

Loading

dstogov commented Apr 24, 2023

dstogov commented Apr 24, 2023

wxue1 commented Apr 25, 2023

dstogov commented Apr 25, 2023

wxue1 commented Apr 25, 2023 •

edited

Loading

wxue1 commented May 12, 2023

dstogov May 15, 2023

wxue1 May 15, 2023

wxue1 commented May 15, 2023

Cacheline demote to improve performance #11101

Cacheline demote to improve performance #11101

Conversation

wxue1 commented Apr 19, 2023

dstogov commented Apr 19, 2023

wxue1 commented Apr 20, 2023 • edited Loading

dstogov commented Apr 20, 2023

wxue1 commented Apr 24, 2023 • edited Loading

dstogov commented Apr 24, 2023

dstogov commented Apr 24, 2023

wxue1 commented Apr 25, 2023

dstogov commented Apr 25, 2023

wxue1 commented Apr 25, 2023 • edited Loading

wxue1 commented May 12, 2023

dstogov May 15, 2023

Choose a reason for hiding this comment

wxue1 May 15, 2023

Choose a reason for hiding this comment

wxue1 commented May 15, 2023

wxue1 commented Apr 20, 2023 •

edited

Loading

wxue1 commented Apr 24, 2023 •

edited

Loading

wxue1 commented Apr 25, 2023 •

edited

Loading