it is a realigner for NGS reads, that doesn't use a lot of ram. Not too sure how it compares to GATK's Local realignment around indels as it is not mentioned. but the authors used reads that were aligned with the popular BWA or BFAST as input. (Bowtie was left out though.)
SRMA was able to improve the ultimate variant calling using a variety of measures on the simulated data from two different popular aligners (BWA and BFAST. These aligners were selected based on their sensitivity to insertions and deletions (BFAST and BWA), since a property of SRMA is that it produces a better consensus around indel positions. The initial alignments from BFAST allow local SRMA re-alignment using the original color sequence and qualities to be assessed as BFAST retains this color space information. This further reduces the bias towards calling the reference allele at SNP positions in ABI SOLiD data, and reduces the false discovery rate of new variants. Thus, local re-alignment is a powerful approach to improving genomic sequencing with next generation sequencing technologies. The alignments to the reference genome were implicitly split into 1Mb regions and processed in parallel on a large computer cluster; the re-alignments from each region were then merged in a hierarchical fashion. This allows for the utilization of multi-core computers, with one re-alignment per core, as well as parallelization across a computer cluster or a cloud. The average peak memory utilization per process was 876Mb (on a single-core), with a maximum peak memory utilization of 1.25GB. On average, each 1Mb region required approximately 2.58 minutes to complete, requiring approximately 86.17 hours total running time for the whole U87MG genome. SRMA also supports re- alignment within user-specified regions for efficiency, so that only regions of interest need to be re-aligned. This is particularly useful for exome-sequencing or targeted re-sequencing data.