High CPU when fog-aliyun is used in CF Cloud Controller

[stephanme]

When switching in a CF landscape with "some usage" the cloud controller from bits-service to fog, we observed a drastic increase in CPU usage of the ruby processes of the Cloud Controller (see attached monitoring screenshot). Such an increase was not observed on AWS, Azure and GCP where we switched from bits to fog as well. The CF usage (# of cf-push operations) should be comparable between the different IaaS.

[stephanme]

[stephanme]

[stephanme]

Another metrics screenshot including memory. CPU could be caused by garbage collection activity. Note that meanwhile CC api VMs run on 4 core machines (before: 2 core). This explains why CPU is now lower. 100% CPU = all cores of the VM are fully used, one ruby process can use only one core (as Ruby works single threaded).

[jochenehret]

We've set up a Cloud Foundry landscape with bits-service on AliCloud to check which load the bits-service can handle. Test setup:

cf-deployment v12.39.0

All CF VM sizes are vm_2cpu_4gb (except for Diego cells which have vm_2cpu_8gb)

Number of "api" nodes: 4 Number of "api_workers": 2 Number of "bits-services": 2

We've pushed a number of test applications which contain larger resource files. The content of the files is randomly generated to make sure that no caching influences the result. Here is the test script:

#!/bin/bash

trap "kill 0" EXIT

APP_DOMAIN="<TODO> insert app domain"
NUMBER_APPS=10
RESOURCE_FILE_SIZE_MB=100

function push_app() (
  while :
  do
    echo "$1 $(date) Pushing cf app test$i ..."
    mkdir -p test$i
    pushd test$i > /dev/null
      touch Staticfile
      echo "Hello, World $i" > index.html
      rm resource$i || true
      dd if=/dev/urandom of=resource$i bs=1M count=$RESOURCE_FILE_SIZE_MB
      cf push test$i --hostname test$i -d $APP_DOMAIN -m 64M > /dev/null
    popd > /dev/null
  done
)
for ((i = 1 ; i <= $NUMBER_APPS ; i++));
do
  echo "Starting subprocess $i"
  push_app $i &
done

echo "Waiting for ctrl c"
wait

To run the script, log on to CF and create a new space. Insert the domain name for the test apps (see TODO above), adapt number of apps and size of resource file and execute the script. It will push CF apps in parallel "for" loops.

Our test results:

NUMBER_APPS	RESOURCE_FILE_SIZE_MB	CC overall requests completed per sec	CC average CPU load	CC average memory
10	10 MB	19	12%	33%
10	100 MB	10	10%	37%
30	10 MB	35	17%	37%
30	100 MB	20	21%	40%
50	10 MB	45	23%	38%
50	100 MB	25	23%	40%

We did not encounter any CC problems during the test runs. All CF pushes were completed successfully. Cloud Controller was not overloaded and the two bits-service instances were also healthy.

We will repeat the test series with CC and fog-aliyun so that we can compare numbers.

[WeiQuan0605]

Repeat the test with Fog-aliyun, here ist the result:

NUMBER_APPS	RESOURCE_FILE_SIZE_MB	CC overall requests completed per sec	CC average CPU load	CC average memory
10	10MB	5	36%	73%
10	100MB	5	39%	80%
30	10MB	(timeout connecting to log server, no log will be shown)/The app is invalid: VCAP::CloudController::BuildCreate::StagingInProgress/cf apps: Server error, status code: 502, error code: 0, message:

[stephanme]

The test results above were measured with fog-aliyun 0.3.10 not yet with the improved 0.3.11 that addresses the memory issues.

[jochenehret]

We've repeated the load tests with fog-aliyun 0.3.15.

Test setup: CF v12.39.0 with 4 api and 2 cc-worker nodes of type "vm_2cpu_4gb". Load test runs #NUMBER_APPS parallel "cf push" processes. Each app has a randomly generated resource file of size RESOURCE_FILE_SIZE_MB.

AliCloud landscape with fog-aliyun 0.3.15:

NUMBER_APPS	RESOURCE_FILE_SIZE_MB	CC overall req/sec	CC avg CPU load	CC avg memory	Worker avg CPU	Worker avg memory
10	10 MB	7	43%	75%	34%	72%
10	100 MB	7	38%	75%	26%	72%
30	10 MB	12	66%	82%	40%	72%
30	100 MB	17	79%	80%	51%	74%
50	10 MB	15	75%	80%	92%	75%

For the last two test runs, the landscape starts to fail (CC becoming unresponsive). In our dashboards we could see that the following CC worker jobs queued up: BlobstoreDelete, DeleteExpiredDropletBlob, DeleteExpiredPackageBlob

To get numbers for comparison, we've run the same tests on a AWS landscape with fog-aws 2.0.1:

NUMBER_APPS	RESOURCE_FILE_SIZE_MB	CC overall req/sec	CC avg CPU load	CC avg memory	Worker avg CPU	Worker avg memory
10	10 MB	17	21%	40%	19%	42%
10	100 MB	12	33%	42%	17%	42%
30	10 MB	43	48%	43%	23%	43%
30	100 MB	30	60%	44%	18%	44%
50	10 MB	55	66%	44%	22%	44%
50	100 MB	50	65%	44%	22%	44%

We can see that the memory usage is almost constant. CPU usage increases, but the landscape was still responsive.

[jochenehret]

We've repeated the performance tests with fog-aliyun 0.3.17 and the same setup as described in the previous comment:

NUMBER_APPS	RESOURCE_FILE_SIZE_MB	CC overall req/sec	CC avg CPU load	CC avg memory	Worker avg CPU	Worker avg memory
10	10 MB	17	18%	45%	16%	52%
10	100 MB	11	18%	47%	15%	52%
30	10 MB	11	39%	46%	17%	52%
30	100 MB	27	26%	45%	25%	70%
50	10 MB	58	51%	46%	18%	69%
50	100 MB	50	41%	46%	17%	69%

Worker memory load is a little higher than on AWS, but the other numbers are comparable or even better.

[xiaozhu36]

@jochenehret Thanks for your feedback. We will continue to improve it performance and the next version will resolve "Worker memory load is a little higher than on AWS" issue.