GCP Dataproc - Slow read speed from GCS

2
votes

I have a GCP dataproc cluster where I'm running a job. The input of the job is a folder where there are 200 part files. Each part file is approx 1.2 GB big.

My job is just map operations

val df = spark.read.parquet("gs://bucket/data/src/....")
df.withColumn("a", lit("b")).write.save("gs://bucket/data/dest/...")

The property parquet.block.size is set to 128 MB which means that each part file will be read 10 times during the job.

I enabled the bucket access logging and looked at the stats and I was surprised to see that each part file is getting access whopping 85 times. I can see that there are only 10 requests which send the actual data other requests are either sending 0 bytes in return or some very small amount.

I do understand that reading a big parquet file in splits is standard Spark behavior. Also there must be some metadata exchange requests as well but 8X calls is something very strange. Also if I take a look at amount of data transferred and time taken it looks like that data is getting transferred at 100 MB/mins speed which is very very slow for google's internal data transfer (from GCS to dataproc). I am attaching a CSV with bytes, time taken, url for one part file.

Has anybody experienced such behavior with dataproc? Is there an explanation for so many requests to the file and such slow transfer rates.

As a side note both bucket and dataproc cluster are in same region. There are 50 workers with n1-standard-16 machines.

enter image description here

Since I could not attach the file I'm pasting the formatted contents here.

| sc_bytes  | time_taken_micros | cs_uri                                                                                                                                  | 
|-----------|-------------------|-----------------------------------------------------------------------------------------------------------------------------------------| 
| 0         | 21000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 17000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 22000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 709922    | 164000            | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 709922    | 86000             | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 709922    | 173000            | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 18000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 8         | 47000             | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 17000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 18000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 17000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 8         | 51000             | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 17000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 18000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 12000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 17000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 8         | 103000            | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 709922    | 98000             | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 8         | 42000             | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 18000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 18000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 17000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 709922    | 88000             | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 8         | 42000             | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 18000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 18000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 20000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 20000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 8         | 40000             | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 17000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 20000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 15000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 19000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 18000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 17000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 143092175 | 63484000          | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 16000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 17000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 19000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 32000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 16000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 19000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 18000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 14000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 137585202 | 66010000          | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 136726977 | 66732000          | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 176684024 | 101921000         | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 32000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 17000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 709922    | 113000            | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 16000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 23000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 16000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 17000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 19000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 134187229 | 64401000          | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 135450987 | 73632000          | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 24000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 21000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 15000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 17000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 15000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 27000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 709922    | 106000            | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 137020002 | 66333000          | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 17000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 17000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 24000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 8         | 41000             | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 25000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 16000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 8         | 39000             | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 20000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 709922    | 135000            | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 16000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 0         | 19000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 709922    | 126000            | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 8         | 41000             | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 0         | 18000             | /storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet                    | 
| 135686216 | 71676000          | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media | 
| 179573683 | 90877000          | /download/storage/v1/b/spark-ml-tkrt/o/Dfs%2F/massiveDf%2Fpart-00184-35441be5-85ca-4b21-85bd-cb99f9aa3093-c000.snappy.parquet?alt=media |
1
apache-sparkgoogle-cloud-platformgoogle-cloud-dataproc
Were you able to find a fix or an optimization for this? - surjikal

1 Answers

2
votes

Relatively large number of GCS metadata requests (URLs without ?alt=media parameter in your table) is expected in this case. Job driver performs metadata requests to list files and get their sizes to generate splits, after that for each split workers perform multiple metadata requests to check if files exist, get their size, etc. I think that this seeming inefficiency stems from the fact that Spark uses HDFS interface to access GCS and because HDFS requests have much lower latency than GCS, I don't think that whole Hadoop/Spark stack was heavily optimized to reduce number of HDFS requests.

To address this issue, on Spark level, you may want to enable metadata caching with spark.sql.parquet.cacheMetadata=true property.

On GCS connector level, to reduce number of GCS metadata requests you can enable metadata cache with fs.gs.performance.cache.enable=true property (with spark.hadoop. prefix for Spark), but it can introduce some metadata staleness.

Also, to take advantage of latest improvements in GCS connector (including reduced number of GCS metadata requests and support for random reads) you may want to update it in your cluster to latest version or use Dataproc 1.3 that has it pre-installed.

Regarding read speed, you may want to allocate more worker tasks per each VM which will increase read speed by increasing number of simultaneous reads.

Also, you may want to check if read speed is limited by write speed for your workload, by removing write to the GCS at the end entirely or replacing it with with write to HDFS or some computation instead.