Since XML often contains a large number of elements with the same name (and type), the lookup of existing names is successful many times. For these "allocations" of already added names, no synchronization is necessary. A second, synchronized lookup ensures to add every name only once.
The resizing of the int and String arrays is probably a big time consumer. Also when one thread needs to resize an array, all the following threads need to wait until this thread has allocated a new, bigger array and has copied the old values to their new destination.
At least while the allocation of the new array, the following threads could still add to the old one, when the resizing is triggered before the old array is really full. The triggering thread would still be slowed down, but some of the later thread would not have to wait.
Also the copying of the part before the trigger could be done without the need to sync with the following threads and the synchronized block would become even shorter.
The resizing process could also be placed into a separate "resizing" thread. Any allocation thread would then only be slowed down while copying the running part of the array (added since the pre-allocation trigger) in the worst case.
The actual gain in performance of this approach is difficult to measure and depends strongly on the concrete usage scenario.
For a quick&easy test of the concurrent allocation of XML names I have simply extended the tests with 16 threads, each allocating all the names.
Protocol SAXON2 SAXON5 STRARRAY COSMOS4 [ms] [kB] [ms] [kB] [ms] [kB] [ms] [kB] ALLOCATE 4250 2623 1122 2623 948 11387 343 2761 READ_PREFIX 95 0 106 0 57 0 57 -8906 READ_URI 81 0 83 0 57 0 57 0 READ_LOCALNAME 87 0 84 0 52 0 52 0 READ_QNAME 905 6235 918 6235 136 6235 84 6235 READ_CLARKNAME 161 2671 175 2671 N/A N/A N/A N/A GET_URI_CODE 827 0 782 0 N/A N/A N/A N/A GET_NS_CODE 1512 0 1407 0 N/A N/A N/A N/A GET_NS_CODE2 65 0 85 0 N/A N/A N/A N/A GET_PREFIX_CODE 604 0 554 0 N/A N/A N/A N/A COMPARE_FP 15 0 15 0 78 0 89 0 COMPARE_NAME 31 0 77 0 82 0 82 0 COMPARE_QNAME 99 0 108 0 89 0 89 0While this is not really a realistic scenario, it shows the improvement of the double-check approach.
Why does Saxon5 need double the time of Saxon2 to compare local names?
While the improvement by replacing the name pool for LivCos (Cosmos4) seems only minor (concurrent access is minimal), the implementation is simpler and thread safe.
A Saxon 220.127.116.11 version, patched with a Saxon5 implementation, shows a gain in performance of about 5% when used in LivCos.