102 lines
3.3 KiB
Diff
102 lines
3.3 KiB
Diff
From 6309460981655bd0899bc5ca65af123c2e42b638 Mon Sep 17 00:00:00 2001
|
||
From: glx <ganli2012@gmail.com>
|
||
Date: Tue, 29 Oct 2024 15:19:06 +0800
|
||
Subject: [PATCH 14/20] Update Details.md
|
||
|
||
---
|
||
docs/Details.md | 80 ++++++++++++++++++++++++++++++++++++++++++++++++-
|
||
1 file changed, 79 insertions(+), 1 deletion(-)
|
||
|
||
diff --git a/docs/Details.md b/docs/Details.md
|
||
index db0b407..14ca694 100644
|
||
--- a/docs/Details.md
|
||
+++ b/docs/Details.md
|
||
@@ -261,4 +261,82 @@ libkperf提供了采集子线程的能力。如果想要在上面场景中获取
|
||
attr.includeNewFork = 1;
|
||
```
|
||
然后，通过PmuRead获取到的PmuData，便能包含子线程计数信息了。
|
||
-注意，该功能是针对Counting模式，因为Sampling和SPE Sampling本身就会采集子线程的数据。
|
||
\ No newline at end of file
|
||
+注意，该功能是针对Counting模式，因为Sampling和SPE Sampling本身就会采集子线程的数据。
|
||
+
|
||
+### 采集DDRC带宽
|
||
+基于uncore事件可以计算DDRC的访存带宽，不同硬件平台有不同的计算方式。
|
||
+鲲鹏芯片上的访存带宽公式可以参考openeuler kernel的tools/perf/pmu-events/arch/arm64/hisilicon/hip09/sys/uncore-ddrc.json：
|
||
+```json
|
||
+ {
|
||
+ "MetricExpr": "flux_wr * 32 / duration_time",
|
||
+ "BriefDescription": "Average bandwidth of DDRC memory write(Byte/s)",
|
||
+ "Compat": "0x00000030",
|
||
+ "MetricGroup": "DDRC",
|
||
+ "MetricName": "ddrc_bw_write",
|
||
+ "Unit": "hisi_sccl,ddrc"
|
||
+ },
|
||
+ {
|
||
+ "MetricExpr": "flux_rd * 32 / duration_time",
|
||
+ "BriefDescription": "Average bandwidth of DDRC memory read(Byte/s)",
|
||
+ "Compat": "0x00000030",
|
||
+ "MetricGroup": "DDRC",
|
||
+ "MetricName": "ddrc_bw_read",
|
||
+ "Unit": "hisi_sccl,ddrc"
|
||
+ },
|
||
+```
|
||
+
|
||
+根据公式，采集flux_wr和flux_rd事件，用于计算带宽：
|
||
+```c++
|
||
+ // 采集hisi_scclX_ddrc设备下的flux_rd和flux_wr，
|
||
+ // 具体设备名称因硬件而异，可以在/sys/devices/下查询。
|
||
+ vector<char *> evts = {
|
||
+ "hisi_sccl1_ddrc/flux_rd/",
|
||
+ "hisi_sccl3_ddrc/flux_rd/",
|
||
+ "hisi_sccl5_ddrc/flux_rd/",
|
||
+ "hisi_sccl7_ddrc/flux_rd/",
|
||
+ "hisi_sccl1_ddrc/flux_wr/",
|
||
+ "hisi_sccl3_ddrc/flux_wr/",
|
||
+ "hisi_sccl5_ddrc/flux_wr/",
|
||
+ "hisi_sccl7_ddrc/flux_wr/"
|
||
+ };
|
||
+
|
||
+ PmuAttr attr = {0};
|
||
+ attr.evtList = evts.data();
|
||
+ attr.numEvt = evts.size();
|
||
+
|
||
+ int pd = PmuOpen(COUNTING, &attr);
|
||
+ if (pd == -1) {
|
||
+ cout << Perror() << "\n";
|
||
+ return;
|
||
+ }
|
||
+
|
||
+ PmuEnable(pd);
|
||
+ for (int i=0;i<60;++i) {
|
||
+ sleep(1);
|
||
+ PmuData *data = nullptr;
|
||
+ int len = PmuRead(pd, &data);
|
||
+ // 有8个uncore事件，所以data的长度等于8.
|
||
+ // 前4个是4个numa的read带宽，后4个是4个numa的write带宽。
|
||
+ for (int j=0;j<4;++j) {
|
||
+ printf("read bandwidth: %f M/s\n", (float)data[j].count*32/1024/1024);
|
||
+ }
|
||
+ for (int j=4;j<8;++j) {
|
||
+ printf("write bandwidth: %f M/s\n", (float)data[j].count*32/1024/1024);
|
||
+ }
|
||
+ PmuDataFree(data);
|
||
+ }
|
||
+ PmuDisable(pd);
|
||
+ PmuClose(pd);
|
||
+```
|
||
+
|
||
+执行上述代码，输出的结果类似如下：
|
||
+```
|
||
+read bandwidth: 17.32 M/s
|
||
+read bandwidth: 5.43 M/s
|
||
+read bandwidth: 2.83 M/s
|
||
+read bandwidth: 4.09 M/s
|
||
+write bandwidth: 4.35 M/s
|
||
+write bandwidth: 2.29 M/s
|
||
+write bandwidth: 0.84 M/s
|
||
+write bandwidth: 0.97 M/s
|
||
+```
|
||
--
|
||
2.43.0
|
||
|