[#642]feat(server): better default options for shuffle server

[advancedxy]

What changes were proposed in this pull request?

1. rss.server.buffer.capacity uses JVM heap size * ratio(0.6) by default
2. rss.server.read.buffer.capacity uses JVM heap size * ratio(0.2) by default
3. rss.server.disk.capacity uses disk space * ratio(0.9) by default

Why are the changes needed?

Fix: #642

Does this PR introduce any user-facing change?

Yes. Three new configurations are introduced, users can specify ratio values
for buffer, read buffer and disk capacity

How was this patch tested?

New UTs.

[codecov-commenter]

Codecov Report

Merging #662 (589cd7e) into master (c70d154) will increase coverage by 2.36%.
The diff coverage is 100.00%.

@@             Coverage Diff              @@
##             master     #662      +/-   ##
============================================
+ Coverage     60.67%   63.04%   +2.36%     
- Complexity     1802     1807       +5     
============================================
  Files           216      202      -14     
  Lines         12458    10523    -1935     
  Branches       1052     1054       +2     
============================================
- Hits           7559     6634     -925     
+ Misses         4494     3543     -951     
+ Partials        405      346      -59     

Impacted Files	Coverage Δ
...a/org/apache/uniffle/server/ShuffleServerConf.java	99.37% <100.00%> (+0.03%)	⬆️
...he/uniffle/server/buffer/ShuffleBufferManager.java	83.50% <100.00%> (+0.75%)	⬆️
...he/uniffle/server/storage/LocalStorageManager.java	87.70% <100.00%> (+1.26%)	⬆️
...rg/apache/uniffle/storage/common/LocalStorage.java	55.41% <100.00%> (+4.09%)	⬆️
.../java/org/apache/uniffle/server/ShuffleServer.java	62.06% <0.00%> (-1.98%)	⬇️
...pache/hadoop/mapreduce/task/reduce/RssFetcher.java	90.76% <0.00%> (-1.54%)	⬇️
deploy/kubernetes/operator/pkg/utils/rss.go
...rnetes/operator/pkg/webhook/inspector/inspector.go
deploy/kubernetes/operator/pkg/utils/config.go
...pkg/controller/sync/shuffleserver/shuffleserver.go
... and 13 more

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[jerqi]

Our memory is equal to rss.server.buffer.capacity + rss.server.read.buffer.capacity + extra memory used for program.

[advancedxy]

Our memory is equal to rss.server.buffer.capacity + rss.server.read.buffer.capacity + extra memory used for program.

If I understand the shuffle server code correctly, read buffer and buffer capacity have some overlapping region, So it's possible rss.server.buffer.capacity + rss.server.read.buffer.capacity > 1.0.

The current default value may a bit larger consider the meta memory required for shuffle server.

[jerqi]

Our memory is equal to rss.server.buffer.capacity + rss.server.read.buffer.capacity + extra memory used for program.

If I understand the shuffle server code correctly, read buffer and buffer capacity have some overlapping region, So it's possible rss.server.buffer.capacity + rss.server.read.buffer.capacity > 1.0.

The current default value may a bit larger consider the meta memory required for shuffle server.

They shouldn't have overlap.

[advancedxy]

Our memory is equal to rss.server.buffer.capacity + rss.server.read.buffer.capacity + extra memory used for program.

If I understand the shuffle server code correctly, read buffer and buffer capacity have some overlapping region, So it's possible rss.server.buffer.capacity + rss.server.read.buffer.capacity > 1.0.
The current default value may a bit larger consider the meta memory required for shuffle server.

They shouldn't have overlap.

IIUC, the buffer caches data in memory, when reading with memory shuffle, the data size is calculated both for buffer and read buffer, these two regions should be considered overlapped region?

[zuston]

Our memory is equal to rss.server.buffer.capacity + rss.server.read.buffer.capacity + extra memory used for program.

From the point of view of production environment, the reserved system memory should be the half of xmx. If not, the full gc will occur when the server meets the high pressure.

[advancedxy]

Our memory is equal to rss.server.buffer.capacity + rss.server.read.buffer.capacity + extra memory used for program.

From the point of view of production environment, the reserved system memory should be the half of xmx. If not, the full gc will occur when the server meets the high pressure.

Ouch, that hurts....

[jerqi]

Our memory is equal to rss.server.buffer.capacity + rss.server.read.buffer.capacity + extra memory used for program.

If I understand the shuffle server code correctly, read buffer and buffer capacity have some overlapping region, So it's possible rss.server.buffer.capacity + rss.server.read.buffer.capacity > 1.0.
The current default value may a bit larger consider the meta memory required for shuffle server.

They shouldn't have overlap.

IIUC, the buffer caches data in memory, when reading with memory shuffle, the data size is calculated both for buffer and read buffer, these two regions should be considered overlapped region?

No, they will copy memory. They shouldn't overlap.

[advancedxy]

No, they will copy memory. They shouldn't overlap.

Oh, I saw the code... This seems inefficient, let's reconsider this when doing off-heap memory management.

[advancedxy]

Our memory is equal to rss.server.buffer.capacity + rss.server.read.buffer.capacity + extra memory used for program.

From the point of view of production environment, the reserved system memory should be the half of xmx. If not, the full gc will occur when the server meets the high pressure.

@zuston do you have any suggestion ratio for buffer and read buffer then.

[zuston]

@zuston do you have any suggestion ratio for buffer and read buffer then.

For us, read buffer size is the half of buffer capacity size. But from the dashboard, the read buffer size could be set to smaller.

[advancedxy]

@zuston do you have any suggestion ratio for buffer and read buffer then.

For us, read buffer size is the half of buffer capacity size. But from the dashboard, the read buffer size could be set to smaller.

@zuston do you have any suggestion ratio for buffer and read buffer then.

For us, read buffer size is the half of buffer capacity size. But from the dashboard, the read buffer size could be set to smaller.

I just changed the default value of buffer capacity and read capacity ratio to 0.6 and 0.2, WDYT? @zuston @jerqi

[jerqi]

LGTM

[advancedxy]

Thanks, @zuston and @jerqi. I'm merging this.