From the lab / Export performance

Non-native means slow? reclaimDB vs Oracle Data Pump

How reclaimDB’s datafile-to-CSV export compares with Data Pump on a shared NFS setup.

After the recent export optimizations, I wanted to see how reclaimDB compared with Oracle Data Pump. I used eight tables with just over 12 GiB of data and ran both tools on the same host.

reclaimDB reads the Oracle datafiles directly and writes CSV. Data Pump exports through the running database. I measured how long each took and checked the row counts for every table.

reclaimDB finished in 43.28 seconds, compared with 47.13 seconds for Data Pump — 8.2% less elapsed time in this test.

Median of four measured runs per tool, with parallelism set to four. All 7,255,149 rows were accounted for.

Test data

Data Pump wrote a data-only dump set. reclaimDB read the original Oracle datafiles and wrote RFC 4180 CSV. The output formats differ, so this comparison measures the time to export the table data, not a complete export-and-import cycle.

The tables contain NUMBER, DATE and VARCHAR2 columns, including NULLs. Table sizes and row widths vary, adding up to 12.146 GiB of logical column values. I used synthetic text made from hash-derived strings repeated within each row. There are no LOBs, and neither the tables nor the exports are compressed.

Each table has its own datafile. All eight tablespaces are read-only.

Storage and host

Both tools ran on the same Linux host. The Oracle datafiles, Data Pump dumps and reclaimDB CSV files all used the same network-mounted NFS share.

Environment used for every measured run
Host6 logical CPUs, about 23 GiB RAM
DatabaseOracle 19c Enterprise Edition 19.32
StorageShared NFSv3 over TCP
NFS client16 connections, 1 MiB read/write requests
Source and outputSame NFS share for both tools

Oracle remained running during both exports.

How I measured the exports

I set parallelism to four for both tools and ran them one at a time in a reserved lab window. Each tool had one warm-up run, excluded from the timing results. I then measured four pairs of exports, switching which tool went first in each pair: Data Pump first, then reclaimDB first, and the same sequence again. I recorded host activity and storage-network traffic throughout; normal Oracle background processes remained active.

Before each export, I synchronized pending writes and cleared the Linux client’s filesystem cache. I left Oracle’s SGA and the NFS server’s cache alone, so only the client filesystem cache was reset between runs.

Four measured runs per tool, parallelism 4
ExportMedianMin–maxSource MiB/sOutput GiB
Oracle Data Pump → dump47.13 s45.76–48.45 s263.912.227
reclaimDB → CSV43.28 s42.81–44.05 s287.412.307

Elapsed time covers process launch through successful exit, including initialization. Throughput uses the same logical source byte count for both tools. Including a final synchronization of the output files gives medians of 47.16 s for Data Pump and 43.44 s for reclaimDB.

reclaimDB’s schema metadata was captured before the timed runs, taking 0.97 s. It includes segment locations so the exporter can read the relevant extents. No source datafile copy was needed.

Checking the row counts

For each table, I checked the Oracle source count against the rows reported in the successful Data Pump log and the CSV record count, excluding the header. I used Data Pump’s logged counts without reimporting the dump.

Matching row counts in all measured and warm-up exports
TableOracle sourceexpdp logCSV records
CUSTOMERS1,048,5761,048,5761,048,576
ORDERS1,048,5761,048,5761,048,576
ORDER_ITEMS1,048,5761,048,5761,048,576
INVOICES873,814873,814873,814
PAYMENTS786,432786,432786,432
EVENTS838,861838,861838,861
DOCUMENT_TEXTS786,432786,432786,432
AUDIT_EVENTS823,882823,882823,882
Total7,255,1497,255,1497,255,149

All benchmark CSV exports also matched the previously verified reference files byte for byte. I kept the full logs and validation records separately.

Using the wizard

I start reclaimdb, choose Datafiles (.dbf) local, point it to the files and select the tables. Then I choose CSV, set an output directory and confirm the export. Here, “local” means file paths available on the host; those paths point to the mounted NFS share.

This recording shows how to use the wizard; it is a separate export, not one of the timed comparison runs. The wizard selected six workers and its default excel-de CSV format (semicolon-separated). The benchmark used four workers and RFC 4180 CSV. That is why the settings and elapsed time in the recording differ from the table above.

Press play to see the wizard. Playback runs at 0.7× speed; long idle waits are shortened.

Result

On this NFS setup, reclaimDB finished about four seconds ahead of Data Pump. All eight tables had matching row counts. That is the result for these intact tables without LOBs; other workloads and storage configurations may behave differently.

← All posts