diff --git a/docs/benchmark_netty_case_report.md b/docs/benchmark_netty_case_report.md new file mode 100644 index 0000000000..7d60b5fc67 --- /dev/null +++ b/docs/benchmark_netty_case_report.md @@ -0,0 +1,141 @@ + + +## Environment + +### Software + +Uniffle 0.9.0, Hadoop 2.8.5, Spark 3.3.1 + +### Hardware + +#### Uniffle Cluster + +| Cluster Type | Memory | CPU Cores | Disk Configuration for Every Shuffle Server | Max IO Read/Write Speed | Quantity | Network Bandwidth | +|--------------|--------|-----------|---------------------------------------------|-------------------------|---------------------------------------|-------------------| +| HDD | 250G | 96 | 10 * 4T HDD | 150MB/s | 2 * Coordinator + 10 * Shuffle Server | 25GB/s | +| SSD | 250G | 96 | 1 * 6T NVME | 3GB/s | 2 * Coordinator + 10 * Shuffle Server | 25GB/s | + +#### Hadoop Yarn Cluster + +2 * ResourceManager + 750 * NodeManager, every machine 12 * 4T HDD + +## Configuration + +Spark's configuration: + + ```` + spark.speculation false + spark.executor.instances 1400 + spark.executor.cores 2 + spark.executor.memory 20g + spark.executor.memoryOverhead 1024 + spark.shuffle.manager org.apache.spark.shuffle.RssShuffleManager + spark.sql.shuffle.partitions 20000 + spark.sql.files.maxPartitionBytes 107374182 + spark.rss.storage.type MEMORY_LOCALFILE + spark.rss.writer.buffer.spill.size 1g + spark.rss.writer.buffer.size 16m + spark.rss.client.send.size.limit 32m + spark.rss.client.rpc.maxAttempts 50 + spark.rss.resubmit.stage false + # Enable Netty mode + spark.rss.client.type GRPC_NETTY + spark.rss.client.netty.io.mode EPOLL + ```` + +Shuffle Server's configuration: + + ```` + rss.storage.type MEMORY_LOCALFILE + rss.server.buffer.capacity 140g + rss.server.read.buffer.capacity 20g + rss.rpc.executor.size 1000 + # Enable Netty mode + rss.rpc.server.type GRPC_NETTY + rss.server.netty.epoll.enable true + rss.server.netty.port 17000 + rss.server.netty.connect.backlog 128 + ```` + +## TPC-DS(SF=40000) + +We use [spark-sql-perf](https://github.com/databricks/spark-sql-perf) to generate 10TB data. + +We use the following special SQL to perform stress testing, it mainly focuses on shuffle, with no data skewness, and has +no practical business implications: + +```` +select SUM(IFNULL(CAST(ss_sold_time_sk AS DECIMAL(10, 2)), 0) + IFNULL(CAST(ss_item_sk AS DECIMAL(10, 2)), 0) + IFNULL(CAST(ss_cdemo_sk AS DECIMAL(10, 2)), 0) + IFNULL(CAST(ss_hdemo_sk AS DECIMAL(10, 2)), 0) + IFNULL(CAST(ss_addr_sk AS DECIMAL(10, 2)), 0) + IFNULL(CAST(ss_store_sk AS DECIMAL(10, 2)), 0) + IFNULL(CAST(ss_promo_sk AS DECIMAL(10, 2)), 0) + IFNULL(CAST(ss_ticket_number AS DECIMAL(10, 2)), 0) + IFNULL(CAST(ss_quantity AS DECIMAL(10, 2)), 0) + IFNULL(ss_wholesale_cost, 0) + IFNULL(ss_list_price, 0) + IFNULL(ss_sales_price, 0) + IFNULL(ss_ext_discount_amt, 0) + IFNULL(ss_ext_sales_price, 0) + IFNULL(ss_ext_wholesale_cost, 0) + IFNULL(ss_ext_list_price, 0) + IFNULL(ss_ext_tax, 0) + IFNULL(ss_coupon_amt, 0) + IFNULL(ss_net_paid, 0) + IFNULL(ss_net_paid_inc_tax, 0) + IFNULL(ss_net_profit, 0)) as sum_all_fields from (select * from (select s.*,c.* from (select *,floor(rand(123)*82857000) as sr from store_sales) s join (select*,floor(rand(123)*82857000)as cr from customer) c on s.sr=c.cr) sc DISTRIBUTE BY sc.ss_customer_sk,sc.ss_item_sk) +```` + +## Read-Write Performance + +Total: Read 10.7TiB, Write 6.4TiB + +| Concurrent Tasks | Type | Single Shuffle Server Write Speed | Single Shuffle Server Read Speed | Tasks Total Time | E2E Time | Netty(SSD) Speedup | Netty(SSD) Total Task Time Reduction | Notes | +|------------------|---------------|-----------------------------------|----------------------------------|------------------|----------|--------------------|--------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| +| 1400 | Netty(SSD) | 0.93GB/s | 1.56GB/s | 268.7h | 12min | - | - | | +| | gRPC(SSD) | 0.75GB/s | 1.25GB/s | 330.4h | 15min | 123.02% | 18.67% | | +| | Netty(HDD) | 0.24GB/s | 0.4GB/s | 1024.4h | 46min | 381.12% | 73.77% | | +| | Spark ESS | 0.5GB/s | 0.82GB/s | 525.5h | 23min | 195.56% | 48.88% | | +| | Vanilla Spark | - | - | __*Failed*__ | - | - | - | | +| 2800 | Netty(SSD) | 1.02GB/s | 1.70GB/s | 450.7h | 11min | - | - | | +| | gRPC(SSD) | 0.86GB/s | 1.44GB/s | 566.4h | 13min | 125.64% | 20.42% | | +| | Netty(HDD) | 0.24GB/s | 0.4GB/s | 2009.9h | 46min | 445.83% | 77.6% | | +| | Spark ESS | 0.5GB/s | 0.68GB/s | 672.3h | 23min | 149.19% | 32.96% | | +| | Vanilla Spark | - | - | __*Failed*__ | - | - | - | | +| 5600 | Netty(SSD) | 1.02GB/s | 1.70GB/s | 896.2h | 11min | - | - | | +| | gRPC(SSD) | 0.80GB/s | 1.34GB/s | 1145.1h | 14min | 127.74% | 21.72% | | +| | Netty(HDD) | 0.22GB/s | 0.36GB/s | 4671.3h | 52min | 520.98% | 80.8% | | +| | Spark ESS | - | - | __*Failed*__ | - | - | - | | +| | Vanilla Spark | - | - | __*Failed*__ | - | - | - | | +| 11200 | Netty(SSD) | 0.86GB/s | 1.44GB/s | 1783.1h | 13min | - | - | | +| | gRPC(SSD) | 0.62GB/s | 1.04GB/s | 2028.2h | 15min | 113.74% | 12.08% | At a concurrency of 11,200, the Shuffle Server becomes very unstable in gRPC mode compared to Netty mode, with higher memory usage and CPU load. It is highly susceptible to encountering OOM issues and is not recommended for use. | +| | Netty(HDD) | 0.20GB/s | 0.34GB/s | 8716.5h | 54min | 488.61% | 79.5% | | +| | Spark ESS | - | - | __*Failed*__ | - | - | - | | +| | Vanilla Spark | - | - | __*Failed*__ | - | - | - | | + +Note: + +1. Read and write operations are essentially happening simultaneously. +2. The calculation formula for `Netty(SSD) Total Task Time Reduction` is as follows: + +```` +Netty(SSD) Total Task Time Reduction = (Tasks Total Time - Tasks Total Time( Netty(SSD) )) / Tasks Total Time * 100% +```` + +3. The calculation formula for `Netty(SSD) Speedup` is as follows: + +```` +Netty(SSD) Speedup = Tasks Total Time / Tasks Total Time( Netty(SSD) ) * 100% +```` + +## Conclusion + +We can draw the following conclusions: + +1. At 1400 concurrency, Vanilla Spark is already incapable of successfully completing tasks, and at 5600 concurrency, + Spark + ESS also fails to complete tasks. However, whether it is HDD or SSD, and whether it is gRPC mode or Netty mode, + Uniffle can all run normally. **Uniffle can significantly improve job stability in high-pressure scenarios**. +2. When comparing using SSDs, **Netty mode brings about a 20% of total task time reduction compared to gRPC mode**. +3. When comparing with Netty mode turned on, **SSD brings about an 80% of total task time reduction compared to HDD**. +4. **Above 11200 concurrency, it is not recommended to use gRPC mode**, as gRPC mode will cause the machine's load + to be much higher than Netty mode, and the Shuffle Server's process will consume more memory on the machine. + Also, it is highly susceptible to encountering OOM issues. + See https://github.com/apache/incubator-uniffle/issues/1651 for more details. \ No newline at end of file